feat(provider): add aimlapi.com model provider

[Lookoff-AIMLAPI]

What this PR does / why we need it?

Adds aimlapi.com (AI/ML API) as a first-class model provider.

MaxKB can already reach AI/ML API today by pointing the built-in OpenAI provider at
https://api.aimlapi.com/v1, but the operator has to type every model name by hand and the
provider shows up as "OpenAI". This PR adds a dedicated provider directory following the
regolo shape (the last externally contributed OpenAI-compatible provider, 1Panel-dev#3041), covering
four model types: LLM, IMAGE (vision), EMBEDDING and TTI (image generation).

Summary of your change

apps/models_provider/impl/aimlapi_model_provider/ — credential + model class per type,
provider class, icon, plus the ModelProvideConstants registration, the
ui/.../provider-data.ts entry and the three .po files (zh_CN / zh_Hant / en_US).

Implementation details that are specific to this API, all of them found by probing it today:

• Request params are built by omitting unset keys, never by sending null. Null tolerance
  differs per model on this API: google/gemini-2.5-flash accepts null everywhere, while
  openai/gpt-4o-mini and deepseek/deepseek-chat return 400 for temperature, top_p,
  seed and tools set to null. tools: null is the worst case — an agent loop that clears
  tools between turns succeeds on turn 1 and 400s on turn 2. filter_optional_params() in
  const.py drops None values on top of the base MaxKBBaseModel.filter_optional_params,
  and test_chat_model_omits_unset_params_instead_of_sending_null locks it in.
• Attribution headers are scoped to the request origin. HTTP-Referer / X-Title identify
  MaxKB (the calling app); X-AIMLAPI-* are the channel headers. They are only attached when
  the configured API URL is api.aimlapi.com, so changing the URL to another vendor or to a
  proxy does not leak them, and caller-supplied headers win on a key clash. The shared constant
  is never mutated. Covered by three tests, including a regex assertion on the partner id —
  a malformed id is not rejected by the API, it just silently loses the attribution.
• Embeddings use the openai SDK with plain strings (same shape as
  openai_model_provider/model/embedding.py) rather than LangChain's OpenAIEmbeddings.
  LangChain tokenises the input into token-id arrays first, and this API's /v1/embeddings
  only accepts strings — it answers 400 details[].path = "input" for the array form. This is
  the one thing a copy of the regolo embedding class would have got wrong.
• TTI size/quality default to auto, which means "do not send the parameter". Image
  models disagree about the allowed values: openai/gpt-image-1 accepts only
  low|medium|high for quality and 400s on standard, which is what the other providers
  in this repo send by default.
• TTI credential validation does not call GET /v1/models. That endpoint is public on this
  API and returns 200 for any key, including a bogus one, so it cannot validate a key. The
  check posts to /v1/images/generations with a placeholder model name instead: an invalid key
  returns 401 and a valid one returns 404, so the key is really validated and no image is
  generated (and nothing is billed). Verified both ways.
• The "output max tokens" tooltip no longer promises a cost bound. On this API max_tokens
  does not reliably bound reasoning tokens — zhipu/glm-5.2 returned 849 tokens against a 120
  cap with finish_reason: "stop". The new string is translated in zh_CN and zh_Hant.

The last commit, chore(aimlapi): fork-only placement — do not send upstream, only moves the
provider to the front of the two hand-ordered lists. It is meant to be dropped before this is
offered upstream. MaxKB has no "recommended"/featured badge for providers, so none was invented.

Model list

Model names are hardcoded per provider in this repo, so every id shipped here was checked
against GET https://api.aimlapi.com/v1/models?include=all (id or alias — checking ids
alone is not enough) and, for the ids in the list below, called live. 30 distinct ids, 0
failures on the catalog check. Users can still type any other model name by hand.

Two notes for whoever refreshes this list later:

• Prefer the dotted spelling: anthropic/claude-sonnet-4.5 is valid, -4-5 is not.
• An id can be both a model and an alias of another model: a request for
  deepseek/deepseek-chat is answered by deepseek-v4-flash (the response model field says
  so). The list therefore ships deepseek/deepseek-v4-flash, whose echo matches what was asked
  for. x-ai/grok-4-6 echoes grok-4.6, same model, different spelling.

Verification

• npm run build (type-check + vite build) in ui/: exit 0.
• python apps/manage.py test <all 12 apps>:
  baseline on a pristine checkout 14 tests / 5 pass / 9 errors, after this change
  21 tests / 12 pass / 9 errors. The 9 errors are identical before and after: the test
  loader imports each app's models package as a test module and Django rejects it
  (doesn't declare an explicit app_label). They are unrelated to this change.
• ruff check on every added file: clean.
• msgfmt --check on the three .po files: clean; the new strings were verified to render
  under zh-hans, zh-hant and en.
• One real inference call through the new provider code path (get_provider(...) .get_model(...), not a mock and not curl), plus tool calling, vision, embeddings and image
  generation:

provider name    : aimlapi.com
provider id      : model_aimlapi_provider
model types      : ['LLM', 'EMBEDDING', 'IMAGE', 'TTI']

[LLM] content    : MaxKB 是一个提供高效数据存储与管理的云服务平台。
[LLM] usage      : {'completion_tokens': 17, 'prompt_tokens': 18, 'total_tokens': 35, ...}
[LLM] headers    : {'HTTP-Referer': 'https://github.com/1Panel-dev/MaxKB', 'X-Title': 'MaxKB',
                    'X-AIMLAPI-Partner-ID': 'part_maxkb', 'X-AIMLAPI-Source': 'agent/maxkb'}
[TOOL] tool_calls: [{'name': 'get_weather', 'args': {'city': '北京'},
                     'id': 'call_6WEz7la7PsqOFjsAj5NAdaUB', 'type': 'tool_call'}]
[IMAGE] content  : Red.
[EMBEDDING] dim  : 1536 head: [-0.01503, -0.01166, -0.01041]
[TTI] params     : {'n': 1}
[TTI] check_auth : True
[TTI] urls       : ['https://s3.aimlapi.com/files/file-01a0652e-e79a-7175-8b46-ec...']
[CREDENTIAL] LLM : True

Credential validation was also exercised with a deliberately wrong key: LLM and TTI both
return False (401 from the API), so a bad key is not silently accepted.

Not verified

• No STT / TTS / RERANKER / TTV / ITV models are registered. This API serves speech-to-text
  and video generation through asynchronous submit endpoints that do not match MaxKB's
  synchronous STT/TTS/TTV classes, so wiring them up needs its own change.
• The provider icon had to be added with git add -f: .gitignore:189 ignores
  apps/models_provider/impl/*/icon/, yet every existing provider icon is tracked.
• Only the model ids listed in the provider file were called live; the rest of the ~350 chat
  models on this API were verified against the catalog only.

Please indicate you've done the following:

• Made sure tests are passing and test coverage is added if needed.
• Made sure commit message follow the rule of Conventional Commits specification.
• Considered the docs impact and opened a new docs issue or PR with docs changes if needed.