Support multi-video SLP files

[talmo]

Right now DREEM training assumes that SLP files contain a single video.

Our data loader should do a lot of validation about the expected inputs.

In particular, it should:

1. Treat contiguously tracked frames within the same video as a clip, and not form a batch for training with large gaps in between consecutive frames (e.g., when you only have frames 1-100 and then 1000-1100)
2. Be aware that there may be multiple videos in the same SLP file, so we shouldn't form a batch across different videos.