fix: use original segment length for vad_end time mapping

[Acelogic]

Summary

• Fixes incorrect VAD timestamp mapping when overlap samples are appended to segments
• vad_end was calculated using the overlap-extended segment_length, but orig_end reflects the original VAD boundary without overlap, creating a ratio mismatch in the time interpolation
• Compute original_segment_length before overlap is added and use it for vad_end; the full segment_length (with overlap) is still used for the audio buffer copy and offset advancement

Details

When whisper_full_with_state processes VAD segments, non-final segments get overlap_samples appended (line 6705 on master) for smoother model transitions. The bug is that segment.vad_end was set to samples_to_cs(offset + segment_length) where segment_length includes this overlap, while segment.orig_end is the raw VAD boundary without overlap.

This mismatch means the linear interpolation at lines 6740-6743:

int64_t orig_time = segment.orig_start + (vad_elapsed * orig_total) / vad_total;

uses a vad_total that is too large relative to orig_total, causing all timestamps within the segment to be shifted earlier than they should be.

The fix moves the boundary clamping before overlap addition and computes original_segment_length from the un-extended boundaries. Only vad_end changes; the memcpy and offset advancement still use the overlap-inclusive segment_length so the audio buffer is unaffected.

Fixes #3683

[danbev]

I believe this is was resolved by #3711, but please let us know if that is not the case.