perf(whisper): reuse encoded audio features for word timestamps

[unohee]

Summary

• reuse DecodingResult.audio_features during word-timestamp alignment
• avoid running the Whisper audio encoder a second time for every decoded window
• preserve the existing mel-based alignment path for direct callers that do not provide encoded features

Why

transcribe() already computes and retains encoded audio features in each DecodingResult. find_alignment() discarded those features and called model.encoder(mel) again before the cross-attention/DTW pass. Passing the existing features removes that duplicate encoder invocation without changing decoding or alignment math.

M1 Max benchmark

Model: mlx-community/whisper-large-v3-turbo, FP16, Korean meeting audio, word_timestamps=True, condition_on_previous_text=False.

Audio	mlx-whisper 0.4.3	This PR	Speedup
65 s, warm p50	3.958 s	2.243 s	1.76x
4 min, warm	10.388 s	6.844 s	1.52x
71 min, single run	188.880 s	130.257 s	1.45x

Parity evidence

On the deterministic 4-minute run, all of the following matched exactly:

• complete text
• 88 segment token arrays and start/end timestamps
• 419 word strings and start/end timestamps
• all word probabilities

The long-form default temperature tuple can sample stochastically after a fallback. Re-running an observed fallback region with the same mx.random.seed() produced exact segment, token, timestamp, and probability parity across 19 segments and 53 words.

Scope

This PR intentionally does not include the separate batched-decoding experiment. Fixed-window batching changed long-form window boundaries and failed the output-parity gate.