Add ability to filter bad proof steps from the data loader

[jasonrute]

Add options to the data loader to filter out:

• tactics with None args
• tactics with term args (reduced to a single identifier)
  • has a tactic arg if tactic_text != interm_tactic_text
• tactics which are no ops
  • see Issue #9

[fidel-schaposnik]

I've now implemented the first two items here. However, note that I've leaned on doing this at the TF-GNN level instead of the loader level, because I think we should stick to the useful conceptual separation:

• The loader limits itself to exposing data from the capnp dataset
• The various modules (TF2, TFGNN, etc) do whatever they like with this data

Eventually, I'd like to clean up the loader from other stuff that is cluttering it and not being used (e.g. train/validation splits), but for now at least I think we should not add here higher-level functionality.

Anyway, maybe we can close this issue and keep #9 as a reminder of what is left?