Reduce peak memory usage by ~362 MB for 4-min audio

- Remove debug members (last_power_spectrum, log1p_input_cpp) from
  MelSpectrogram, saving ~61 MB that persisted across the full pipeline
- Release mono_audio immediately after resampling to free ~42 MB early
- Convert float→double per chunk inside resample_audio instead of
  pre-allocating a full-size double copy, saving ~84 MB
- Replace O(T×N) Viterbi backtrack matrix with checkpointed approach
  (K=sqrt(T) interval): peak BT memory drops from ~181 MB to ~6 MB
  while producing bit-identical output (verified by diff on test audio)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0155xf2TorZZUxwf5ZDfixvg