Migrate the kaldi-fbank ASR models to AudioProcessor
Migrates parakeet, nemotron_asr_streaming, cohere_asr, lasr, qwen3_asr,
speech_to_text and seamless_m4t — the models whose frontend is a kaldi-style
log-mel filterbank (povey/hann windowing, per-frame preemphasis, dithering).
`NemotronAsrStreamingAudioProcessor` subclasses `ParakeetAudioProcessor`
directly, which lets `modular_nemotron_asr_streaming.py` drop its duplicated
feature-extraction block.
13 auto entries move to the new classes, covering the sub-model-types that share
these frontends (parakeet_ctc/encoder/rnnt/tdt, lasr_ctc/encoder,
nemotron3_5_asr, nemotron_asr_streaming_encoder, seamless_m4t_v2).