add mms asr model

[mm65x]

Context

MMS (Massively Multilingual Speech) from Meta supports 1000+ languages through language-specific adapter layers on a shared wav2vec2 backbone. It fills a gap in language coverage that no other model in mlx-audio currently provides.

Description

This adds MMS ASR support by building on the existing wav2vec2 implementation. The main additions are a CTC decoding head, adapter layer support in the encoder, and automatic loading of language-specific adapter weights.

The model loads the base wav2vec2 weights from model.safetensors, then overlays language-specific adapter layers and CTC head from adapter.{lang}.safetensors. Audio is normalized to zero mean and unit variance before processing, matching the HF feature extractor behavior.

Changes in the codebase

• mlx_audio/stt/models/mms/mms.py: CTC model wrapping wav2vec2, greedy CTC decoding, adapter and vocab loading
• mlx_audio/stt/models/mms/tests/test_mms.py: 14 unit tests
• mlx_audio/stt/models/wav2vec/wav2vec.py: added Wav2Vec2AttnAdapterLayer and optional adapter support in stable layer norm encoder layers
• mlx_audio/stt/utils.py: register "mms" in MODEL_REMAPPING

Changes outside the codebase

None.

Additional information

• Tested with facebook/mms-1b-fl102, output matches HF transformers reference
• Adapter layers loaded in post_load_hook from separate safetensors files
• Supports multi-language vocab.json format (language-keyed)
• Adapter only added to StableLayerNorm encoder variant, matching HF implementation

Checklist

• Tests added/updated
• Documentation updated
• Issue referenced - STT expansion

[lucasnewman]

@mm65x Looks great! If you can resolve the conflicts we can merge.

[mm65x]

rebased on main, conflicts resolved. also removed the unnecessary from_pretrained deprecation wrapper from moonshine while at it

[Blaizzy]

LGTM, thanks!