Add OpenAI-compatible POST /v1/audio/transcriptions route

[sumgup0]

This adds an OpenAI-compatible transcription route on the existing local
service so any standard OpenAI client/frontend can use TransClip as its
transcription backend (base_url=http://127.0.0.1:8765/v1; the SDK appends
/audio/transcriptions).

Contract (deliberate compat subset): multipart file (formats = what
the bundled soundfile/libsndfile decodes — WAV/FLAC/OGG/mp3 on libsndfile
1.2.2; m4a/webm unsupported, no ffmpeg); model/language/prompt/
temperature accepted and ignored (single configured backend serves, the
local-OpenAI-server norm); response_format json (default, {"text"})
or text (text/plain); anything else 400. OpenAI error JSON shape on this
route only. 25 MB cap (26_214_400, OpenAI's own limit) enforced against
Content-Length BEFORE the body is read — which also gives the server the
request-size bound it previously lacked on this path. Streaming,
verbose_json/srt/vtt, and /v1/audio/translations are explicitly out.

Dictation path untouched: the route is a self-contained branch in
do_POST (after _guard(), before the JSON dispatch and its exception
wrapper) with its own writer and its own error handling — the internal
{"error": str, "debug_capture_dir": ...} shape is unreachable from it
(pinned by test). The only shared-surface changes: a thin
InferenceEngine.transcribe_raw() (raw ASR — no cleanup/voice-mode/
history/debug-capture: OpenAI semantics) and authorization added to the
CORS allow-headers (browser OpenAI clients preflight it; grants nothing —
the loopback-origin guard still gates everything).

Safety details: undecodable audio → 400 via soundfile pre-validation in
the route, with a static message (libsndfile's exception text embeds the
server temp path — never echoed; same discipline for the 500 path, both
pinned by tests). Uploaded-filename suffix sanitized before the tempfile
call. A module-level lock serializes THIS route's backend calls; note the
pre-existing service posture (no inference lock across /transcribe /
dictation) is unchanged — flagging rather than fixing in a thin PR.

The over-limit path drains the request body (bounded, in chunks) before
sending the 413, so a real oversize upload receives a clean status instead
of a connection reset that an SDK would retry.

Tests: 29 unit (multipart parser incl. UTF-8/RFC 2231 filenames, binary

• trailing-CRLF + stray-Content-Transfer-Encoding round-trips, sanitizer,
  handler with fakes) + 10 socket round trips (json/text/400s, a real
  oversize upload → clean 413, 411, 500-no-leak, 403, preflight) via the
  existing serve_test_engine harness + a new multipart helper. Full suite:
  490 OK (5 skipped) at this branch.

One pre-existing quirk noticed, not touched (exists at ba25cb0): the
ruff format non-compliance in server.py's _guard/_cors blocks (and
_worker_loop). Also, on Windows the pre-existing _guard early-reject
path (403 before the body is read) can occasionally reset the connection
before a client with a body in flight reads the response — the new route's
own reject paths drain first to avoid this, but I left the shared _guard
untouched in a thin PR; happy to apply the same drain there if you'd like.