Re-encode plain WAV sources with an unspecified channel layout

[rubas-agent]

Before

A plain mono WAV from ffmpeg -f lavfi -i sine=d=1:sample_rate=44100 a.wav has channel_layout=unknown in ffprobe. On main, every call that resamples or re-encodes it fails:

Exmpeg.extract_audio("a.wav", "x.mp3")      # also .m4a, .opus, .flac with sample_rate: 16_000
Exmpeg.transcode("a.wav", "y.m4a", audio_codec: "aac")
{:error, %Exmpeg.Error{reason: :io_error, message: "AVERROR(-1668179713): `Input changed`"}}

The new tests, run without the Rust fix:

1) test transcode re-encodes a mono 44.1 kHz WAV to aac in m4a
   message: "AVERROR(-1668179713): `Input changed`"
2) test extract_audio re-encodes a mono 44.1 kHz WAV to mp3 (same error)
3) test extract_audio resamples a mono 44.1 kHz WAV to 48 kHz opus (same error)
Result: 61/64 passed

Problem

ffmpeg writes mono and stereo WAV files without a channel mask. The decoder keeps the unspecified channel order. swr_init sets its input to the default layout for the channel count. Then swr_convert_frame sees the unspecified layout of each decoded frame as a changed input and rejects it. An AAC source works, because its layout is set.

After

All calls above return {:ok, _}. task test:integration: 107 passed (104 existing tests and 3 new ones). The WAV-to-WAV passthrough test still passes. task check exits 0.

	main	this PR
Integration tests	104	107
WAV re-encode calls that fail	all	none

Details

Cause. native/exmpeg_native/src/extract_audio.rs:extract_audio and native/exmpeg_native/src/transcode.rs:build_audio_pipeline open the decoder with AV_CHANNEL_ORDER_UNSPEC from codecpar. The resampler normalises its input layout, so the frames no longer match it.

Design. The new helper audio::default_unspecified_layout gives an unspecified decoder layout the default layout for its channel count (AVChannelLayout::from_nb_channels), before decoder.open. The ffmpeg CLI does the same in guess_input_channel_layout (fftools/ffmpeg_demux.c). Both pipelines call the one helper. After it, the decoder layout is never unspecified, so I removed the unspecified branch of the passthrough check in extract_audio. A WAV with the default layout still matches the target and skips the resampler. No unsafe.

Tests. Three integration tests use the existing mono 44.1 kHz WAV fixture:

• extract_audio to .mp3 keeps 44 100 Hz and 1 channel.
• extract_audio to .opus resamples to 48 000 Hz.
• transcode to .m4a with audio_codec: "aac" re-encodes 1 stream.

All three fail on main with the error above and pass with the fix.