v0.55.4

• Audio decoded from sources with more than two channels now downmixes to stereo
  and keeps the correct frame count and array shape. Explicit channels values are
  limited to 1 or 2, which matches the operations supported by Audio.
• Saving audio now writes the PCM width declared by its metadata. Files with 8-bit,
  16-bit, and 32-bit samples keep their duration and signal level.
• Audio slicing rejects starts beyond the track, and sliding analysis rejects
  window and overlap values that cannot advance.
• Video slices now normalize negative bounds and keep audio aligned. Slice steps
  other than 1 raise ValueError; use a speed operation to change playback rate.
• Changes to AI integrations, dependencies, or default models now require manual
  model verification before release.