Whisper: start each clip_timestamps clip at its own start time

[shw1606]

Problem

The clip loop checks only where a clip ends, never where it starts:

for seek_clip_start, seek_clip_end in seek_clips:
    while seek < seek_clip_end:

seek carries over from the previous clip, so the loop decodes the gap between
two clips as if you had asked for it. Pass three clips and you get back one
span running from the first clip's start to the last clip's end.

This file is a port of openai/whisper, which clamps seek on every
iteration:

while clip_idx < len(seek_clips):
    seek_clip_start, seek_clip_end = seek_clips[clip_idx]
    if seek < seek_clip_start:
        seek = seek_clip_start

The comment above that loop reads "A later commit should turn this into a
simpler nested loop", and it describes the nested form this file uses. The port
took the shape and left the clamp behind.

Fix

Clamp seek to the clip start on entering each clip, the nested-loop
equivalent of what upstream does. The line does nothing unless you pass
clip_timestamps.

Test

test_transcribe_clip_timestamps requests clip_timestamps="0,1,3,4" on the
bundled ls_test.flac and checks which windows the loop decodes. Setting
no_speech_threshold=None keeps every visited window in the output, so the
test can assert that each segment's seek lands inside a requested clip and
that the loop reaches both clips.

Window starts, in seconds:

	visited
before	[0.0, 1.0], and 1.0s sits inside the excluded 1s-3s gap
after	[0.0, 3.0], the two clip starts

On main the second window decodes the gap and returns "Good soul openly
sorted the boat and sh...", text the caller asked to skip. The test fails on
main, passes with this change, and runs in 1.2s on the existing tiny
fixture.

End-to-end effect

141s of Korean speech: five utterances, 21s of speech, 30s gaps of digital
silence and inaudible noise between them. mlx-community/whisper-turbo on an
M1 Pro (16GB), macOS 26.5, mlx 0.32.2, clips from silero-vad. Median of
three trials with warm weights:

	main	this PR
wall clock	27.4s	4.2s
segments returned (5 expected)	30	5
segments outside the requested clips	26	0
consecutive duplicate segments	26	0
segment-start error vs ground truth	8.46s	0.23s

Drop clip_timestamps and the two builds agree: 4.26s vs 4.30s median and the
same output, as a no-op path should behave.

The slowdown needs an explanation, since skipping audio ought to save time.
segment_size is min(N_FRAMES, content_frames - seek, seek_clip_end - seek),
so once the loop walks into a gap, the next clip boundary truncates the window
to a sliver. Whisper decodes a long tail of short near-silent windows instead
of a few full ones, and most of those fall through to temperature fallback and
loop on repeated text. #1285 reports the same shape behind a silero-vad
front end: a 40s file taking "~3-4 seconds" without clip_timestamps and
"~27-28 seconds" with it. That points at this bug as the cause.

Reproduction

The test audio is five say -v Yuna utterances separated by 30s blocks of
digital silence and -50 dBFS noise. I take clips from
silero_vad.get_speech_timestamps, pad them by 0.3s, merge any pair less than
1s apart, and hand the result to clip_timestamps. The "outside the requested
clips" row counts segments whose [start, end] no requested clip contains. I
can attach the benchmark scripts.