Minor fragment environment speed improvements + replay fix

[bengioe]

Changes:

• Caches Data instance of graphs
• Compute the remove_node_mask more efficiently
• Remove the deepcopy in sampling iterator and fix a data overwriting bug
• Data cache of modified graphs is cleared in step with Graph.clear_cache

[julienroyd]

LGTM, I'll just run a training with goal-conditioning to validate that we don't have issues with the replay buffer anymore.

[julienroyd]

Edit: LGTM, the results seem to reproduce previous behavior (with deepcopies)