Expose per-step decoder cross-attention scores during generate() (for streaming / AlignAtt policies)

The Whisper generate() result (WhisperGenerationResult) exposes sequences, scores, logits and no_speech_prob, but not the decoder cross-attention. Attention is only available via the separate align() method, and return_attention exists for the Translator path (see #1716).
This blocks streaming policies like AlignAtt (used by SimulStreaming), which decide token-by-token whether a token is "safe" to emit based on where attention is focused in the audio. Because per-step cross-attention isn't exposed during generate(), projects built on faster-whisper / CTranslate2 can't implement AlignAtt and fall back to the slower torch backend.
Reference: ufal/SimulStreaming#10 — the maintainer confirms faster-whisper can't be used precisely because it doesn't expose attention.
Would it be possible to add a return_attention=True option to Whisper generate() that returns per-step cross-attention scores? The existing align() implementation and the Translator's return_attention suggest the underlying data and mechanism are already available.
Related: #1716, #1855.

SimulStreaming Issue #10

Start by reading the existing Whisper align() implementation and the Translator path described in #1716, focusing on how return_attention is handled. Trace how Whisper generate() produces WhisperGenerationResult values. Done means generate() accepts return_attention=True and exposes per-step decoder cross-attention scores for streaming policies such as AlignAtt.