Add music generation agent tool

[octo-patch]

Reason: The multimodal tool registry covers image, video, TTS, voice and audio mixing but has no music generation tool.

Changes

• multimodal_core.py: added a generate_music capability alongside the existing voice methods, reusing the current region resolution so both the global and CN /v1/music_generation endpoints are supported, together with the shared Bearer/JSON header helper.
• Request body is built by _minimax_music_payload, which sends model (the only mandatory field) plus prompt, lyrics, output_format, audio_setting.format, is_instrumental, lyrics_optimizer and an explicit non-streaming stream flag. Optional fields are omitted when the caller does not supply them, and aigc_watermark is only attached on the CN region.
• Output-format handling supports both documented forms: hex payloads are decoded to bytes, url payloads are downloaded before the temporary link expires, and both are stored under /storage/audios like the other audio tools. Audio format accepts mp3, wav and pcm; unsupported values are rejected before any request is sent.
• Response parsing is isolated in _extract_minimax_music_audio: it checks base_resp.status_code, only reads data.audio for the completed status, and reports a clear message for the in-progress status and for empty audio.
• New minimax_music.py tool class exposing minimax_music_generation, registered in the planner/ReAct flow tool list, in the multimodal branch of agent_task_runner, and in the UI compact tool-content set so its results render like the sibling audio tools.
• New settings in .env.example: MINIMAX_MUSIC_MODEL (default music-3.0), MINIMAX_MUSIC_OUTPUT_FORMAT, MINIMAX_MUSIC_AUDIO_FORMAT and MINIMAX_MUSIC_AIGC_WATERMARK. No model name or host is hard-coded at the call site; the existing MINIMAX_REGION / MINIMAX_BASE_URL settings continue to select the host.

Checks

• python3 -m pytest tests/test_minimax_music_tool.py tests/test_minimax_voice_tool.py --noconftest -q — 14 passed (10 new tests plus the 4 existing voice tests). --noconftest was needed because tests/conftest.py imports the FastAPI app, whose database dependencies are not installed in this environment.
• python3 -m py_compile on every changed Python module.
• Drove the tool end-to-end against a stubbed HTTP layer to confirm the request reaches /v1/music_generation with the Bearer header, that a completed hex response is decoded to a stored file, that a url response is downloaded and stored, and that an in-progress response fails with a readable message.

New tests cover the tool schema registration, endpoint resolution for both regions, payload defaults and optional fields, the CN-only watermark field, completed / in-progress / error / empty-audio response parsing, the missing API key guard, the prompt-or-lyrics requirement, and format validation.