transformers

d4c6ab25 - Migrate kyutai_speech_to_text: legacy FE becomes a deprecation alias (issue 21)

Migrate kyutai_speech_to_text: legacy FE becomes a deprecation alias (issue 21) KyutaiSpeechToTextAudioProcessor already existed and the auto mapping already pointed at it, but a 177-line KyutaiSpeechToTextFeatureExtractor survived in the modular source, duplicating the processor's padding logic under the legacy name. Removed from the modular file; feature_extraction_kyutai_speech_to_text.py is now a make_legacy_audio_processor_alias shim. Verified the modular converter regenerates the model files without clobbering the shim. The two classes carried different defaults (0.0/0.0 on the extractor, 1.0/2.5 on the processor), but every published checkpoint pins both values, so from_pretrained is unaffected -- and those repos are fingerprinted, so a behaviour change would fail the hub suite. Alias verified: FutureWarning on instantiation, isinstance and issubclass against the processor, and AutoAudioProcessor still resolves kyutai configs (which declare feature_extractor_type: KyutaiSpeechToTextFeatureExtractor) to the processor with config values intact.