Adds ActionIndex and Harmonize use of action-masks

[julienroyd]

Masks were used somewhat ad-hoc here and there. Here we centralise masking in GraphActionCategorical(), specifically:

• renamed masks in GraphActionCategorical to action_mask (to avoid confusion with done-masking, molecule validity masking, goal-conditioning masking, etc.)
• made logits and masks private attributes of GraphActionCategorical
• remove _mask() application from GraphTransformerFragEnvelopeQL and GraphTransformerGFN
• removed masking from GraphSampler (masks applied in property setters of GraphActionCategorical._logits)

We also replaced the action-tuples Tuple[int, int, int] action type, row and column by a named-tuple ActionIndex to the code and typing more intuitive and readable.

[julienroyd]

Can reproduce previous results with goal-conditioning and 4 objectives: