Harmonize use of masks

[bengioe]

Masks are currently used somewhat ad-hoc here and there, in particular:

• GraphSampler explicitly uses them, but it should be abstracted away
• GraphTransformerGFN and derivatives end up applying masks themselves, but this feels like the "wrong" place

Instead, it would be good for GraphActionCategorical to do all the accounting of masks and mask-related code.