Multi-size videos in data pipelines

[talmo]

Support for datasets with videos with different image shapes in the data pipeline

Addresses the longstanding #209.

[talmo]

The approach here should be:

1. Determine target height/width (max of all videos or user-specified)
2. Resize longest side to target height/width
3. Pad (bottom/right) the short side to target width/height

Other notes:

• This metadata will need to be saved in the config so these steps can be reproduced during inference time. This may not be strictly necessary since we can retrace for different sizes as long as each batch is from the same video, but we should test.
• This padding should be aware of/compatible with the padding done to account for model stride (sleap.nn.data.resizing.pad_to_stride/Resizer). Ideally they'd be done in the same step, but this gets complicated for multi-model approaches (topdown) that may require different padding.