whisper: align no_speech fallback logic with upstream

[AdityaBiranje]

Summary

Fixes #1427

decode_with_fallback in mlx_whisper/transcribe.py suppressed fallback
decoding whenever no_speech_prob exceeded no_speech_threshold, regardless
of avg_logprob. Upstream openai/whisper only suppresses fallback when
the segment looks like silence by both measures — high no_speech_prob
and low avg_logprob.

Because the MLX condition was broader, a segment that triggered fallback for
an unrelated reason (e.g. high compression ratio / repetitive output) could
still get marked "silence" and have fallback wrongly cancelled, letting
low-quality decodes through undetected.

Change

whisper/mlx_whisper/transcribe.py: added the missing logprob_threshold
/ avg_logprob check to the "silence" condition, matching upstream:

if (
    no_speech_threshold is not None
    and decode_result.no_speech_prob > no_speech_threshold
    and logprob_threshold is not None
    and decode_result.avg_logprob < logprob_threshold
):
    needs_fallback = False  # silence

I extracted this into a small _needs_fallback helper so the logic could be
unit tested directly, rather than only through the decode_with_fallback
closure.

Testing

• Added a unit test covering the case that previously broke: high
  no_speech_prob with low compression ratio/logprob threshold still
  triggering fallback correctly.
• Ran the existing test suite in whisper/tests/ — all passing.
• Manually verified on a short local audio sample that decode behavior
  matches upstream openai/whisper on the same clip.

Notes

No public API changes. This only affects the internal fallback-suppression
condition and should make behavior more consistent with upstream.