Datamodules that have fixed train/val/test should allow for easy subsampling of train

[calebrob6]

For running SSL experiments similar to those from the SeCo paper (see screenshot below), we need a way to randomly get "10%" of the training data from a particular dataset. This should probably be handled at the datamodule level. Posting an issue because I'm not sure when I'll get to this.

[adamjstewart]

@yichiac was trying to do something like this recently