[Feature] Add v1 STT integration for Whisper models (#143)

## Summary

Wire the STT pipeline (PRs #96, #98, #126, #133, #137) into vLLM's v1
engine so that `vllm serve openai/whisper-small` serves
`/v1/audio/transcriptions` and `/v1/audio/translations` endpoints.

- **model_runner**: STT model loading with caching, dummy KV cache spec
(Whisper self-manages KV), warm-up from model config, greedy decode
loop, audio feature extraction (handles `MultiModalKwargsItem`/UserDict,
bfloat16 torch tensors, shape transpose)
- **platform**: STT auto-detection, tokenizer fallback (only when
unset), disable `async_scheduling`
- **worker**: skip paged attention for STT, nominal memory for
scheduler, `get_supported_tasks` returns `("transcription",)`
- **docs**: add `whisper-large-v3-turbo` to model table

## Test

<img width="1453" height="432" alt="截圖 2026-03-07 中午12 54 23"
src="https://github.com/user-attachments/assets/241d81ee-1437-4ecd-8a93-ea5eca6937ca"
/>
<img width="1436" height="328" alt="截圖 2026-03-07 中午12 55 16"
src="https://github.com/user-attachments/assets/6ff2b61d-bdd0-4023-911d-266296933ae2"
/>

<img width="864" height="193" alt="截圖 2026-03-07 下午1 06 37"
src="https://github.com/user-attachments/assets/f2af98c7-083c-44d9-ba31-4f50f9e62ae6"
/>

## Verification 
### Unit tests
```
pytest tests/test_v1_stt_integration.py tests/test_stt.py tests/test_whisper.py tests/test_transcribe.py -v -m "not slow"
```

### End-to-end
```
pytest tests/test_v1_stt_integration.py -v -m slow
```

### Server smoke test (requires local model)
```
vllm serve /path/to/whisper-small-mlx --port 8000
```

```
curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@test.wav \
  -F model=whisper
```

---------

Signed-off-by: RickyChen / 陳昭儒 <ricky.chen@infinirc.com>

Author: ricky-chaoju

docs/stt.md

@@ -50,6 +50,7 @@ Any OpenAI Whisper checkpoint (HuggingFace or MLX format):
 | Whisper Small | 244M | `openai/whisper-small` |
 | Whisper Medium | 769M | `openai/whisper-medium` |
 | Whisper Large V3 | 1.5B | `openai/whisper-large-v3` |
+| Whisper Large V3 Turbo | 809M | `openai/whisper-large-v3-turbo` |
 
 MLX-format weights (e.g. from `mlx-community`) are also supported.