Fix word-level timestamp overflow in Whisper chunked transcription

[neonwatty]

Summary

• Fixes whisper-large-v3-turbo_timestamped has broken timestamps #1357
• Clamps word-level timestamps to the actual chunk_len to prevent timestamps from exceeding audio duration
• When the model outputs timestamps near the 30s boundary (e.g., 29.98s) for shorter final chunks, timestamps would incorrectly exceed the audio duration

Root Cause

When using chunk_length_s=30 with word-level timestamps (return_timestamps: 'word'), the Whisper model can output timestamps up to ~29.98s (the maximum representable timestamp given time_precision = 30/1500 = 0.02). For a final chunk shorter than 30s, these timestamps would be added to the accumulated time_offset, causing the final timestamps to exceed the actual audio duration.

Solution

Track the actual chunk_len from the stride information and clamp raw token timestamps before adding time_offset. This ensures word-level timestamps never exceed the audio duration while preserving the existing behavior for segment-level timestamps.

Test plan

• Added unit test that simulates the bug case (65s audio with 15s final chunk)
• Verified fix with actual whisper-large-v3-turbo model on HuggingFace GPU
• All existing tests pass
• Build succeeds