[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. Based on v1.8.4.1.

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
• When windowed attention is active, the encoder bypasses flash attention and uses the standard softmax path (Metal flash attention does not support custom F32 masks)
• Flash attention remains enabled for non-BCI models and for the decoder
• Adds proper SOS token (language + transcribe) initialization for BCI models

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 completely unaffected

Context

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 on v1.8.4.1
• Verified locally: 10.4% average WER across 5 BCI test samples (identical to v1.7.6 baseline)

Test results

BCI package (@qvac/bci-whispercpp)

• 4/4 integration tests pass (9/9 assertions)
• 10.4% average WER across 5 neural signal samples — identical output to v1.7.6 baseline

Standard whisper package (@qvac/transcription-whispercpp)

• Integration: 10/10 tests pass (40/40 assertions)
• Unit: 13/13 tests pass (57/57 assertions)
• No regression in standard audio transcription (s16le + f32le formats verified)