[BCI] QVAC-17071 feat: add BCI neural signal support (variable conv1 kernel + windowed attention)

[sharmaraju352]

Summary

Adds two changes to whisper.cpp to support brain-computer interface (BCI) neural signal transcription:

1. Variable conv1 kernel size

• Reads n_audio_conv1_kernel from model hparams (defaults to 3 for standard whisper models)
• Allows BCI models to use a different first convolution kernel size

2. Windowed self-attention for encoder layers

• Adds n_audio_window_size and n_audio_last_window_layer hparams
• When present, encoder self-attention is restricted to a local window for layers up to last_window_layer
• Adds proper SOS token (language + transcribe) initialization for BCI models
• Both flash-attention and standard attention paths are updated

Backward compatibility

Both changes are backward-compatible:

• n_audio_conv1_kernel defaults to 3 (standard whisper behavior)
• n_audio_window_size defaults to 0 and n_audio_last_window_layer defaults to -1, which disables windowed attention entirely
• Standard whisper models are unaffected

Context

These changes are required by the new @qvac/bci-whispercpp addon: tetherto/qvac#1583

Test plan

• Standard whisper transcription still works (no regression)
• BCI model loads and transcribes neural signals correctly
• Verified locally: 10.4% average WER across 5 BCI test samples

Made with Cursor

[sharmaraju352]

Closing in favor of a new PR based on v1.8.4.1 with flash attention fix.