fix: allow flash attention in encoder when DTW is enabled

[Acelogic]

Summary

• DTW token timestamps no longer silently disabled when flash attention is enabled
• Flash attention remains active for encoder self-attention and decoder self-attention
• Only cross-attention falls back to the non-flash path (needed to extract KQ_soft_max weights for DTW)

Details

Previously, whisper_init_with_params_no_state disabled DTW entirely when flash_attn was set:

if (params.flash_attn && params.dtw_token_timestamps) {
    params.dtw_token_timestamps = false; // DTW silently disabled
}

DTW only needs the explicit cross-attention weights (KQ_soft_max) from the decoder, which the flash attention path doesn't produce (it fuses QKV into one operation). The encoder self-attention and decoder self-attention don't interact with DTW at all.

The fix introduces a flash_cross flag (flash_attn && !dtw_token_timestamps) used at two matching locations:

1. Encoder (cross-attention KV storage layout) — determines how K/V are stored in kv_cross
2. Decoder (cross-attention computation) — determines whether to use the fused flash path or the explicit K*Q → softmax → KQV path

Both locations must use the same condition to keep the KV cache layout consistent.

Test plan

• Builds cleanly on macOS with Metal
• Functional testing requires a whisper model with DTW-compatible alignment heads

Fixes #3662