Core data loader

[talmo]

• PyTorch (PTL?) Dataset
• Support loading SLP files using sleap_io
• Be able to return individual LabeledFrames and the corresponding Instances and image in a dictionary of numpy arrays

This should be analogous to LabelsReader in main SLEAP.

Notes:

• Be mindful of tensor ordering! PyTorch expects channel-first ordering (frames, channels, width, height), so we should use that ordering as soon as possible (but check what the augmentation backends prefer in Augmentation pipeline block #2). The idea is to minimize the number of times we do permutations -- more specifically, we want to minimize the number of copies we have to make of the large image arrays. It might be possible to get around this with Tensor Views though see the notes on this:

  Taking a view of a contiguous tensor could potentially produce a non-contiguous tensor. Users should pay additional attention as contiguity might have implicit performance impact.

• Use multiprocessing/concurrency if possible at this step (but might need to happen at the DataLoader level later?)