fix(stt): support tokenizer.json for cohere-transcribe-03-2026 (mlx_audio ≤ 0.4.2 compat)

[guglxni]

Summary

• mlx_audio ≤ 0.4.2 post_load_hook hardcodes tokenizer.model (SentencePiece), but CohereLabs/cohere-transcribe-03-2026 ships only tokenizer.json (HuggingFace fast tokenizer), causing a crash at server startup.
• Adds _CohereAsrTokenizerHF — a drop-in shim over HuggingFace tokenizers implementing the same interface (eos_token_id, build_prompt_tokens, batch_decode) as the existing CohereAsrTokenizer.
• Monkey-patches Model.post_load_hook inside load_model() to probe tokenizer.model first, then fall back to tokenizer.json — same pattern already used for load_buffers_from_checkpoint.

Root cause

# mlx_audio/stt/models/cohere_asr/cohere_asr.py  line 705
tokenizer_path = model_path / "tokenizer.model"   # ← always SentencePiece
model._tokenizer = CohereAsrTokenizer(str(tokenizer_path), ...)
# CohereAsrTokenizer.__init__ calls spm.SentencePieceProcessor().load(path)
# → OSError / RuntimeError because the file doesn't exist

CohereLabs/cohere-transcribe-03-2026 (released March 2026) moved to the HuggingFace tokenizers JSON format; mlx_audio 0.4.2 was not updated to match.

Test plan

• python stt_server.py --model CohereLabs/cohere-transcribe-03-2026 --port 8101 starts without error
• POST /v1/audio/transcriptions returns a transcript for a WAV file
• Models that still ship tokenizer.model (SentencePiece) continue to load via the original path

Notes

The R2 bundle ships mlx_audio 0.2.7 (no cohere_asr module at all). This fix is only exercised on installs that have upgraded to 0.4.2 and are targeting cohere-transcribe-03-2026. A proper upstream fix in mlx_audio would be preferable long-term.

[vercel]

@guglxni is attempting to deploy a commit to the EigenLabs Team on Vercel.

A member of the Team first needs to authorize it.

[guglxni]

@Gajesh2007 could you approve the workflow run for this PR? Thanks!