Trim the AC-3 decoder delay in transport streams - #21

AC-3 and E-AC-3 decode 256 samples (5.3 ms at 48 kHz) late. MP4 and MOV
record that in their edit list, which FFmpeg applies. MPEG-TS, the container
of AVCHD (.MTS/.M2TS), records it nowhere, so FFmpeg's raw output keeps
it, and on Windows and Linux AVCHD audio sat 5.3 ms earlier than on macOS,
where AVFoundation trims the delay.

The ffprobe parser now marks AC-3 and E-AC-3 streams in MPEG-TS with that
256-sample decoder_delay, and the FFmpeg tier starts them that much later
for whole-file analysis, refinement windows and native (export) decoding.
Every other stream keeps a delay of 0.

Exports follow the trimmed timing:

• The trim leaves the container's last 256 samples without audio. Channel
  stems and placement pads fill that gap with silence, as they already fill
  a camera's sub-frame tail, so an audio-only transport stream still exports
  at its container length. Before, the shortfall failed the stem with
  "Cannot decode audio for drift correction." Audio missing anywhere else is
  still rejected.
• Clean-media export of cameras outside QuickTime and MP4 runs FFmpeg on the
  recorder directly. It now skips the recorder's decoder delay too, instead
  of placing its audio 256 samples late.

The fingerprint cache namespace is not bumped here: fingerprints of AC-3 in
MPEG-TS made before this change are 5.3 ms early. This ships after
fix/damaged-audio-packets, which moves the namespace to portable4; if it
were released without it, portable3 would need the bump.

Evidence:

(The camera clips were synchronized against their own audio extracted from
the first video frame; what matters is that both engines now agree.)

Not changed: AAC in MPEG-TS. Its encoder priming is recorded nowhere either,
and the engines disagree in opposite directions (AVFoundation trims 2112
samples, FFmpeg none, an FFmpeg-encoded stream primes 1024), so there is no
reliable correction. AVCHD uses AC-3 or LPCM.

Tests, each failing without its fix:

• ac3_in_transport_streams_starts_on_time places a click encoded at 1.5 s
  exactly, from the head, after a seek, and in a refinement window. Without
  the trim it lands at sample 72256 instead of 72000.
• stems_of_audio_with_a_trimmed_decoder_delay_keep_their_length: a stem of
  a whole audio-only transport stream has the container's length, with the
  click on time.
• ffmpeg_remux_trims_the_recorder_decoder_delay: a clean-media export of an
  MPEG-TS camera places the recorder's click at sample 31200, not 31456.

Validation:

• cargo fmt --all -- --check
• cargo clippy --locked --workspace --all-targets -- -D warnings
• cargo test --locked --workspace: 376 passed, 1 ignored