WAVE_FORMAT_EXTENSIBLE decodes on Python 3.12 but raises on 3.10 (the CI version)

[thorwhalen]

The gap

WAVE_FORMAT_EXTENSIBLE (format tag 0xFFFE) WAV files — which ffmpeg writes routinely, and which any multichannel or >16-bit file tends to use — decode on Python 3.12 and raise on Python 3.10:

python 3.10.13:  decode_wav_bytes(extensible_wav) -> wave.Error: unknown format: 65534
python 3.12.12:  decode_wav_bytes(extensible_wav) -> ([0, 1, -1, 2, -2], 44100)

Same bytes, same recode. CI runs 3.10 only, so this is invisible to the suite.

Cause

Not in recode's own parsing. The chunk walker added in #11 handles these files correctly — it locates the payload at the right offset (verified: _wav_data_chunk returns (68, 10) on both versions). The failure is one line later, in decode_wav_header_bytes, which delegates to stdlib wave:

# 3.10  wave._read_fmt_chunk
if wFormatTag == WAVE_FORMAT_PCM: ...
else: raise Error('unknown format: %r' % (wFormatTag,))

# 3.12  wave._read_fmt_chunk
if wFormatTag != WAVE_FORMAT_PCM and wFormatTag != WAVE_FORMAT_EXTENSIBLE:
    raise Error(...)

stdlib wave gained EXTENSIBLE support in 3.12.

Options

1. Parse the fmt chunk directly. The walker is already there and fmt is a fixed layout of 16/18/40 bytes. This makes behaviour identical on every supported Python and drops the stdlib wave dependency from the read path — wave is a deprecated-adjacent module and its Wave_read is also why the errors were heterogeneous before Fix #4: locate the WAV data chunk instead of inferring where it must be #11.
2. Declare EXTENSIBLE out of scope and say so in the docstring.
3. Raise the supported Python floor to 3.12 (probably not worth it for this alone).

Option 1 is the one that makes recode self-contained on the read path, and the machinery is now mostly in place.

Found by adversarial review while fixing #4.