[FedyaLight]

Stacked on the #16 branch (fix/damaged-audio-packets), which introduced the
extension-named container readers this change configures.

MP3 encoders prime their output: LAME adds 1105 samples of delay at the start
and padding at the end, and records both in its LAME/Info header. FFmpeg
trims them. Align's in-process decoder (Symphonia, on Windows and Linux) did
not, so MP3 audio sat 23 to 25 ms early and synchronized that far off — more
than half a frame at 25 fps.

The delay Symphonia reads from the LAME header (encoder delay plus the
decoder's 529 samples, 1105 for LAME) now becomes the stream's lead, which
whole-file decoding, seeks, refinement windows and native decoding already
skip for MP4 edit lists. Files without a LAME header are unchanged (nothing
records their delay; FFmpeg cannot trim them either).

Only the start is trimmed, deliberately. Symphonia's own gapless mode, tried
first, also trims the end at a frame count that files without a Xing header
merely estimate from their first frames, which dropped the audio after a quiet
stretch, and it underflows when the Xing frame count contradicts the LAME
delay. End padding (under one frame) does not affect synchronization.

Evidence, on the 59-file decoder differential: the three MP3 files (CBR, VBR
with ID3, 44.1 kHz mono) moved from 1.977, 1.977 and 1.975 s to exactly
2.000 s; nothing else changed.

Tests:

• aac_priming_matches_ffmpeg_timing becomes
  encoder_priming_matches_ffmpeg_timing and also covers MP3 (skipped
  without libmp3lame). Without the trim it fails with the click at sample
  73105 instead of 72000, the 1105-sample LAME delay.
• mp3_frame_counts_never_cut_audio: a headerless VBR file (loud, silent,
  loud) decodes as long as FFmpeg's output, and a Xing frame count of 1
  neither panics nor shortens decoding. With Symphonia's gapless mode it
  decodes 94464 of 200448 frames.

The fingerprint cache namespace moved to portable4 in the #16 branch. If
this change ships in a later release than that one, bump it again so cached
MP3 fingerprints are recomputed.

Validation:

• cargo fmt --all -- --check
• cargo clippy --locked --workspace --all-targets -- -D warnings
• cargo test --locked --workspace: 382 passed, 1 ignored