[FedyaLight]

Fixes #16.

On Windows and Linux, Align decodes MP4, MOV and WAV audio with Symphonia.
Its AAC decoder reports a truncated or corrupted packet as an I/O error,
"unexpected end of bitstream". Align treated every such error as fatal, and
one file's failure stopped the whole run, without naming the file.

Reproduced with a camera-original Canon MP4 (CanonMP4_002, AAC-LC) with one
audio packet cut short: 0.3.2 stops with exactly the reported
Decode error: symphonia: unexpected end of bitstream., and this branch
synchronizes the file at the same offset as the intact copy.

What changes for the user

Nothing is dropped that a decoder can read:

• A damaged audio packet becomes silence of its recorded duration. The rest of
  the file is used and keeps its timing, which FFmpeg itself does not do: it
  drops such packets, and sometimes stops the stream.
• Symphonia is chosen only when the first decodable packet does not yield far
  fewer samples (under three quarters) than its container records. HE-AAC,
  whose SBR layer Symphonia lacks, yields half and did not synchronize; it now
  goes to FFmpeg. PCM is not checked, and short final packets pass. M4A is
  decoded like MP4, with its edit list.
• When the in-process decoder (Symphonia, or AVFoundation on macOS) still
  fails on a file during analysis, FFmpeg decodes it again from the start,
  into fresh state, also for a mix of all streams.
• Only a file that neither can read is reported, in a warning that names it,
  and the other files still synchronize. In 0.3.2 such a file stopped the
  whole run. When only some streams of a file can be read, refinement uses the
  stream the analysis kept.
• Cancellation ends the run. So does a missing FFmpeg, except on macOS, where
  Align does not ship FFmpeg and AVFoundation's verdict on a file stands.
• The portable fingerprint cache namespace moves to portable4, so
  fingerprints the previous decoder produced (HE-AAC, misdetected MXF) are not
  reused. The first synchronization after the update analyses files again.

Symphonia 0.5 also ignores the probe hint and scans the first megabyte of any
file for MP3 or ADTS sync words. It detected a camera-original Canon EOS C70
XF-AVC MXF as 32 kHz MPEG audio after reading 4.3 MB, decoded generated MXF
files as 11 kHz MP3 noise, and read whole files (160 MB of 160 MB) when the
reader rejected its match. Containers are now opened by the reader their
extension names: MXF, MTS and R3D go straight to FFmpeg.

The status switches to "Analyzing" when audio analysis starts. Before, it
showed "Reading metadata" until the first file was analysed, which is where
the report saw the error. Analysis progress is counted and published under
one lock, so the completed count no longer goes backwards when workers
finish together.

Decoder differential

59 files carrying the same noise at a known offset: WAV (8 to 64 bit, RF64,
6 channels), AIFF, MP3, AAC-LC/HE-AAC/HE-AACv2/ALAC in M4A, MOV with PCM, AAC
and ALAC, MP4 (plain, faststart, fragmented), MXF, MTS with AC-3, AAC and
LPCM, files with damaged packets, a truncated file, a zeroed region, files
with no audio, plus camera-original Canon C70 MXF, Canon MP4 and Canon 6D
MOV. Each was synchronized against a reference on the portable backend, and
FFmpeg decoded each one independently.

• No file is reported unreadable.
• Every file FFmpeg reads gives the same result as 0.3.2, offsets included,
  except these, which improve:

File	0.3.2	This branch
Canon MP4, one damaged packet	run stopped	synchronized, exact offset
M4A, 5 % of packets damaged (FFmpeg stops after 8 of 20 s)	run stopped	synchronized, exact offset
M4A, about 80 % of packets damaged (FFmpeg reads 1 s)	run stopped	used, unmatched
MP4 with HE-AAC	unmatched	synchronized through FFmpeg

The differential also shows offsets that are wrong in 0.3.2 and unchanged
here, for separate follow-up: MP3 by 23 to 25 ms (encoder delay not
trimmed), HE-AAC by 64 ms (also through FFmpeg), fragmented MP4 and MTS with
AAC by 21 ms (AAC priming), MTS with AC-3 by 5 ms.

Tests

Each fails on main or on the previous revision of this branch for the
intended reason:

• damaged_packet_becomes_silence_of_its_duration
• streams_decoded_shorter_than_their_packets_go_to_ffmpeg
• mp3_frames_inside_other_containers_are_not_decoded_as_audio
• aac_priming_matches_ffmpeg_timing now also covers M4A
• short_final_pcm_packets_still_decode (the last milliseconds of an AIFF)
• in_process_decode_failures_are_retried_with_ffmpeg, for one stream and for
  all streams mixed; only one of the two files fails, so samples kept from the
  failed attempt would shift it
• refinement_uses_the_stream_that_could_be_read
• analysis_progress_never_goes_backwards (deterministic: the listener holds
  odd counts, so overlapping publication would reorder them)
• unreadable_audio_is_reported_and_the_rest_still_synchronizes (the audio
  data is missing, so neither decoder can read it)

The retry and stream-selection tests were also checked by mutation: keeping
the failed attempt's samples, skipping the retry for mixed streams, or the
old stream selection each makes them fail.

Validation

• cargo fmt --all -- --check
• cargo clippy --locked --workspace --all-targets -- -D warnings
• cargo test --locked --workspace: 381 passed, 1 ignored
• Acceptance runner with the default and the portable backend: release and
  preserve-editing pass. The AAF gates were not run (no AAF module or
  pyaaf2 on this machine); this change does not touch AAF.
• Windows was not run locally.

The reporter's files are not available. A damaged packet reproduces the
exact message; the MXF misdetection is a separate defect found on footage
from the same camera family.