Support NWB training packages (PoseTraining)

[talmo]

In ndx-pose 0.2.0 and since rly/ndx-pose#24, NWB now supports "training packages" (PoseTraining) which don't require timeseries and which are better suited to store training data that come from random frames.

See also: rly/ndx-pose#29

We should support this format natively in sleap-io. We don't make a distinction between the two different pose types (inference vs training) since our Labels is flexible enough to incorporate both. This means that we'll want to have some manual control over which format to store things in when we call sio.save_nwb, perhaps with a smart default as a heuristic.

We'll also need to figure out the best way to auto-detect which format to load in sio.load_nwb.

Finally, there's the issue of embedded image data (see also #44).