PYA-1783: Add multichannel support

[MatiasDiBernardo]

Optional multichannel support in RTTM loading

What

The RTTM channel field (column #3) was parsed but discarded. This adds an opt-in way to keep it, so a multichannel recording (telephone, per-headset,source-separated) can expose one reference Annotation per channel.

API

• load_rttm(path, keep_channel=True) → nested {uri: {channel: Annotation}}. Default keep_channel=False is unchanged.
• RTTMLoader returns the annotation of file["channel"] when the ProtocolFile provides one; otherwise it merges all channels (previous behavior).

Key example

Given this RTTM (column #3 is the channel):

SPEAKER conv1 1 0.0 2.0 <NA> <NA> speaker_A <NA> <NA>
SPEAKER conv1 2 1.5 1.0 <NA> <NA> speaker_B <NA> <NA>
SPEAKER conv2 1 0.0 1.0 <NA> <NA> speaker_C <NA> <NA>

Without keep_channel (default — unchanged): all channels merge into one Annotation.

>>> load_rttm("conv.rttm")
{'conv1': <Annotation>, 'conv2': <Annotation>}
>>> load_rttm("conv.rttm")['conv1'].labels()
['speaker_A', 'speaker_B']        # channels 1 and 2 merged

With keep_channel=True — one Annotation per channel:

>>> load_rttm("conv.rttm", keep_channel=True)
{'conv1': {1: <Annotation>, 2: <Annotation>}, 'conv2': {1: <Annotation>}}
>>> load_rttm("conv.rttm", keep_channel=True)['conv1'][2].labels()
['speaker_B']                     # just channel 2

At the protocol level, the same distinction is driven by a channel key on the file:

RTTMLoader("conv.rttm")({"uri": "conv1"})               # -> merged (speaker_A + speaker_B)
RTTMLoader("conv.rttm")({"uri": "conv1", "channel": 2}) # -> speaker_B only

Tests

tests/test_loader.py (+ a multichannel fixture) covers the default/merged path, the keep_channel shape and content, RTTMLoader channel selection and its backward-compatible default, and lazy channel resolution through ProtocolFile. uv run pytest tests is green.

[hbredin]

We'll discuss this offline but I believe this brings a breaking change for RTTMLoader("conv.rttm")({"uri": "conv1", "channel": 2}):

• Before: returns annotations from all channels
• After: only returns for channel 2

I believe the new proposed behavior makes more sense but the ins and outs need to be discussed.