Add multipart audio alignment endpoint

[0xShug0]

Summary

Adds POST /v1/audio/alignments for forced alignment with uploaded audio bytes. This is the upload-based counterpart to /v1/tasks/run for remote or Docker clients that cannot pass a server-local audio path.

Request shape:

curl http://localhost:8080/v1/audio/alignments \
  -F "file=@speech.wav" \
  -F "model=qwen3-align" \
  -F "text=The task has completed successfully."

The endpoint requires file, model, and text; language and busy_timeout_ms are optional. It only accepts WAV uploads for now and requires the selected model to be configured with task=align and mode=offline.

Fixes #350.

Example Response

{
  "text": "Some call me nature. Others call me Mother Nature.",
  "language": "English",
  "words": [
    {
      "word": "Some",
      "start": 0.266667,
      "end": 0.64,
      "start_sample": 6400,
      "end_sample": 15360,
      "confidence": 0
    }
  ],
  "timing": {
    "wall_ms": 123.0,
    "audio_duration_ms": 456.0,
    "rtf": 0.27
  }
}

Validation

Built the debug server and validated multipart behavior with assets/resources/sample.wav.

Case	Response
Qwen3 forced aligner upload	200, words=25, first word includes start, end, start_sample, and end_sample
MMS forced aligner upload	200, words=24, first word includes start, end, start_sample, and end_sample
non-multipart request	400, audio alignment requests must use multipart/form-data
missing file	400, multipart alignment request requires a non-empty 'file' field
empty file	400, multipart alignment request requires a non-empty 'file' field
missing model	400, multipart alignment request requires a 'model' field
missing text	400, multipart alignment request requires a non-empty 'text' field
non-WAV upload	400, only WAV audio uploads are currently supported for alignment
bad busy_timeout_ms	400, multipart busy_timeout_ms field must be an integer
negative busy_timeout_ms	400, busy_timeout_ms must be >= 0 (0 means no client-side bound)
unknown model	400, unknown model id: missing-align-model
wrong task model	400, audio alignment requires a model configured with task=align

[Geequlim]

Tested and accepted.

I ran the Docker image built from this PR and verified the new multipart endpoint with the project fixture assets/resources/a.wav and its known transcript.

The request was sent as multipart/form-data to POST /v1/audio/alignments with file, model=qwen3-align, language=en, and text. The server returned HTTP 200 with all 16 expected words, monotonic timestamps, and consistent sample offsets. The first request loaded the model successfully; a warm request completed in about 23 ms server-side (RTF 0.0039).

The implementation matches the intended scope: dedicated upload route, required-field and WAV validation, in-memory audio decoding, task/mode checks, reuse of the existing alignment runtime, timing/sample-offset response fields, and updated documentation. All GitHub CI checks are passing.

One non-blocking observation: the returned confidence value is 0 for each word, matching the current response example. This does not affect the timestamp acceptance for this feature.

No blocking issues found. Accepted from my side — thank you!

[AnatolyRugalev]

This is a very much needed feature, thank you