Add cross-attention accessor functions for AlignAtt streaming

[danielbodart]

Summary

• Adds whisper_decode_with_state_and_aheads() — same as whisper_decode_with_state but saves alignment head cross-attention data during decode
• Adds whisper_state_get_aheads_cross_qks() — reads the resulting cross-attention tensor from state, copying from GPU/backend to CPU

These enable external callers to implement AlignAtt-style streaming policies by accessing the cross-attention data that is currently only available internally via DTW timestamp computation.

Motivation

Building a streaming speech-to-text system that uses cross-attention analysis (AlignAtt) to decide when to stop decoding and emit partial results. The existing API only exposes DTW timestamps as a post-processing step via whisper_full, but streaming requires access to the raw cross-attention data during manual decode loops.

Details

• whisper_decode_with_state_and_aheads calls whisper_decode_internal with save_aheads=true, matching how whisper_full internally enables attention head saving for DTW
• whisper_state_get_aheads_cross_qks returns a float pointer of shape [n_tokens × n_audio_ctx × n_heads], valid until the next call or whisper_free_state
• Requires dtw_token_timestamps=true in context params (same prerequisite as DTW timestamps)
• No changes to existing functions or data structures — purely additive