Migrate the Gemma/USM-style audio models to AudioProcessor
Migrates gemma3n, gemma4, gemma4_unified and inkling — the models using the
USM-style extended framing, where frames are cut at `win_length + 1` and
reduced back to `win_length` by a per-frame preemphasis. That shape is now
expressed by `StftConfig.frame_extension` and `preemphasis_mode` rather than
by bespoke framing code in each feature extractor.
