Separate P_B trunk & min_len

[bengioe]

This PR makes it possible to have a separate set of parameters for P_B by setting cfg.model.do_separate_p_b. It also introduces cfg.algo.min_len, which defaults to 0 and prevents and agent from stopping before min_len steps have been taken.

Other changes:

• This change required adding a timestep parameter to graph_to_Data across contexts.
• While debugging I added a consider_masks_complete flag to contexts, which indicates when masks computed by a context should be taken as ground truth in terms of which actions are valid. This should help in some aspects in the future (e.g. counting parents, debugging new masks).
• Adds explicit mask computation in SeqTransformerGFN

[bengioe]

Basic experiments on the double trunk model (2pPB) vs shared parameterized (pPB) vs uniform (uPB). The other parameter is sampling_tau