Follow the temperature schedule in batched fallback

[CodeWithBehnam]

What does this PR do?

Closes #5.

• Batched decoding always decoded at 0.0, then retried failing segments once at 1.0, whatever temperature the caller passed:
  • temperature=0.0 still triggered a retry.
  • The default schedule jumped straight to its noisiest step.
  • Retried results were never re-checked.
• decode_batch_with_fallback now walks the same schedule as the one-window path. At each step it re-decodes only the segments that still fail, as one smaller batch. A single temperature means no fallback.
• The pass/fail check (needs_fallback) is shared by both paths instead of being duplicated.
• best_of is still dropped in batched mode, now with a comment saying why: DecodingTask only supports it for a single segment.
• The test stub now numbers windows by their spectrogram content, so a re-decode maps back to the same window.

How was this tested?

• Tested with audio file(s)
• Ran existing tests (pytest): 30 passed
• Tested CLI (vayu audio.mp3)

tests/test_temperature_fallback.py covers:

• a single temperature means no fallback (batch sizes 1 and 3)
• only the failing window is re-decoded at 0.2, then 0.4
• the last result is kept when the schedule runs out
• silent segments don't fall back
• the sequential path still walks the schedule

The three batched tests fail on main.

claude-review will fail as on #12 (the repository's CLAUDE_CODE_OAUTH_TOKEN secret, see this comment).

🤖 Generated with Claude Code

https://claude.ai/code/session_01TMXYMqgLAykqRApmbRfpTA

---

Generated by Claude Code