Add backward sampling

[bengioe]

This PR adds proper backward sampling routines, specifically to sample from parameterized backward policies or to sample uniformly (from masks).

This PR also:

• Changes how stem attachments are defined in the fragment environment using a saner "src_"/"dst_" format
• Includes Recursive decompose routine for fragment env #96 to showcase offline training
• adds an extra Queue object for in-main-process calls to an MP wrapped model
• changes GraphActionCategorical._compute_batchwise_max to use finfo().min instead of the batch min to compute the argmax -- this avoids some edge case errors.
• numpy-fies graph_to_Data in the fragment environment

Includes @diamondspark as an author since the PR is based on his implementation.

[hadim]

See a minor comment about the use of if.