First person view support

[SudeepDasari]

All of RoboNet data thus far is collected in third person view. It'd be nice to add in support in the for loading first person data. Specifically, that means:

• Adding metadata to each record for whether each camera view is in first person or not
• Allowing users to filter on this metadata (i.e only train on first person data)
• Add support in RoboNetDataset for loading records with mixes of first and third person views. Specifically, users should be able to only load the camera views which are from the specified "frame of reference" while ignoring the others.