Handle audio data URL MIME variants

[fus3r]

Summary

Audio data URL prefix stripping only matched word-only MIME subtypes. Valid data URLs such as data:audio/x-wav;base64,..., data:audio/vnd.wave;base64,..., and data:audio/wav;codec=pcm;base64,... were left untouched.

That made audio parsing try to base64-decode the full data URL rather than the payload, so valid audio bytes failed with a base64 or audio-format error. This affected both AudioChunk.to_openai() and tokenizer/audio parsing through Audio.from_base64() / Audio.from_audio_chunk(). The same bytes worked when the MIME subtype was simply audio/wav.

This broadens the audio data URL prefix match to allow MIME subtype punctuation and media-type parameters before ;base64,, and makes Audio.from_base64() reuse the same helper as AudioChunk.to_openai(). The parameter match avoids consuming semicolon separators, so malformed inputs with repeated semicolons do not trigger pathological backtracking. The payload and detected format are unchanged.

Testing

• pytest tests/test_converters.py tests/test_audio.py -q
• ruff check src/mistral_common/protocol/instruct/chunk.py src/mistral_common/tokens/tokenizers/audio.py tests/test_converters.py tests/test_audio.py
• ruff format src/mistral_common/protocol/instruct/chunk.py src/mistral_common/tokens/tokenizers/audio.py tests/test_converters.py tests/test_audio.py --check
• mypy src tests scripts
• pytest tests/ --ignore=tests/integrations -q
• pytest --doctest-modules ./src -q
• mkdocs build --strict
• pytest tests/integrations/ -q

No related issue. This follows the same audio conversion surface as #245.