Make flash_attention effective for the Whisper encoder

[nilsreck]

Problem

Setting flash_attention=true on a Whisper model has no effect today, for two independent reasons:

1. WhisperEncoderLayer constructs its TransformerEncoderLayer without forwarding model.use_flash_attention(), so the parameter always falls back to its default (false) and the encoder never takes the flash path (src/layers/whisper.cc).
2. FlashMultiHeadAttention constructs ops::FlashAttention without is_causal, which defaults to true (src/layers/flash_attention.cc). That is correct for decoder self-attention — currently the only in-tree user of the flash path — but wrong for any encoder, which must attend bidirectionally. This latent bug blocks enabling flash attention for every encoder, not just Whisper's.

Fix

• Pass is_causal = _is_decoder when constructing ops::FlashAttention. Behavior-neutral for decoders; makes the flash path correct for encoder self-attention.
• Forward model.use_flash_attention() from WhisperEncoderLayer to its TransformerEncoderLayer.

The scope is deliberately limited to Whisper: TransformerEncoder (the generic one) also reads model.use_flash_attention() without forwarding it to its layers, but I only validated the Whisper encoder numerically, so I left the generic path unchanged. With the is_causal fix in place, wiring it up becomes a safe follow-up.

This does not touch wheel packaging — it only affects source builds with -DWITH_FLASH_ATTN=ON.

Validation

Measured with a WITH_FLASH_ATTN=ON source build (based on a93c2bf) on one NVIDIA RTX PRO 6000 (Blackwell), Whisper large-v3-turbo, float16:

• Encoder forward pass: 358 ms → 116 ms (3.1×)
• Cosine similarity of encoder output vs. the non-flash encoder: 0.999667
• Pairwise WER on our internal test set: −0.026

Happy to share the comparison script if useful.