Fix torchcodec audio decoding to respect 'num_channels'

[AsymptotaX]

Fixes torchcodec audio decoding when num_channels is set on Audio.

Before this change, AudioDecoder["array"] reduced multi-channel audio to mono by averaging channels, so the requested channel behavior was not respected.

With this PR:

• multi-channel decoded arrays are preserved by default;
• mono output is returned only when num_channels == 1 is explicitly requested.

Previous behavior

ds_stereo = ds.cast_column("audio", Audio(num_channels={...})) - None, 1, 2

Original file: '(16000, 2)' - stereo ✓
'Audio(num_channels=None)': '(16000,)' - mono ✗
'Audio(num_channels=2)': '(16000,)' - mono ✗
'Audio(num_channels=1)': '(16000,)' - mono ✓

New behavior

'num_channels=None' preserves the original number of channels from the source file.
'num_channels=2' preserves/converts to stereo output with shape '(2, num_samples)'.
'num_channels=1' downmixes to mono with shape '(num_samples,)'.

Results

Original file shape (via soundfile): (16000, 2)
HF datasets shape with num_channels=None: (2, 16000)
HF datasets shape with num_channels=1: (16000,)
HF datasets shape with num_channels=2: (2, 16000)

Fixes #8005.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[lhoestq]

Hi ! I believe ["array"] is used a lot in transformers and has been expecting mono for a long time - from before we switched to torchcodec. So we might need to keep mono as default for ["array"] unless num_channels is specified explicitly. Would it be ok for you ?

I see in your tests that you expect stereo even if num_channels is not specified in Audio()