Add audio transcriptions proxy support

[EthenJ]

Summary

• add /v1/audio/transcriptions protocol proxy routing
• forward multipart audio bodies unchanged to the active relay endpoint
• normalize common relay base URL shapes for audio transcription endpoints
• cover route matching, URL normalization, and multipart forwarding with tests

Test plan

• cargo fmt --all
• cargo test -p codex-plus-core --test protocol_proxy audio_transcriptions -- --nocapture
• cargo check -p codex-plus-core
• npm run vite:build in apps/codex-plus-manager
• npm run check in apps/codex-plus-manager
• cargo check --workspace after building frontend assets
• local E2E: POST /v1/audio/transcriptions through Codex++ gateway using gpt-4o-mini-transcribe

[BigPizzaV3]

感谢贡献，音频转录代理这个方向有价值。审查后目前有一个需要先处理的阻塞问题，暂不合并：

handle_helper_connection() 仍通过 read_http_request() 读取请求，而该解析器只根据 Content-Length 判断请求体是否完整，不处理 Transfer-Encoding: chunked。音频 multipart 上传比普通 JSON 请求更可能由 SDK/流式文件使用 chunked 传输；这种情况下当前代码会在读完请求头后就结束读取，导致上游收到空或截断的音频体。

当前 audio_transcriptions_proxy_forwards_multipart_body 测试直接调用 open_audio_transcriptions_proxy_request()，绕过了本地 helper 的 TCP 请求解析，因此无法发现这个问题；测试体也只是短 ASCII 文本，没有验证二进制字节是否原样保留。

请补充：

1. helper 对 chunked multipart 的正确读取/解码，或者在该入口明确限制并返回可识别错误，不能静默截断；
2. 从本地 TCP helper 入口开始的完整回归测试，覆盖分段到达的请求头/请求体；
3. 包含 0x00、非 UTF-8 字节的 multipart 文件内容，并断言上游收到的字节完全一致；
4. 请求大小上限边界测试，并明确当前 32 MiB 限制是否符合预期。

修复后可以继续审查。

[EthenJ]

Thanks @BigPizzaV3 for the detailed review and for catching the helper-layer request parsing gap.

I pushed an incremental fix that:

• decodes Transfer-Encoding: chunked request bodies before forwarding;
• preserves multipart audio data as raw bytes, including 0x00 and non-UTF-8 bytes;
• rejects ambiguous or unsupported framing with HTTP 400;
• returns HTTP 413 when the decoded body exceeds 32 MiB;
• adds TCP-level regression coverage starting at the local helper;
• covers fragmented headers, irregular chunks, chunk extensions, trailers, binary payloads, incomplete bodies, and exact size boundaries;
• uses incremental chunk scanning to avoid repeatedly rescanning large uploads.

Validation

• Standard multipart E2E: HTTP 200, exact transcript.
• Raw HTTP/1.1 chunked E2E: HTTP 200, valid transcript.
• Chunked helper body and upstream body were both 187,338 bytes.
• Three concurrent multipart requests returned HTTP 200.
• Core tests: 131 passed.
• Protocol proxy tests: 46 passed.
• Frontend build and TypeScript checks passed.
• cargo check --workspace passed.
• Release build passed.

The chunked E2E request was manually encoded over TCP with 95 irregular chunks, fragmented headers and chunk-size lines, binary WAV data, and a trailer header.