Regression Transformer - facilitate scaffold decoration

[jannisborn]

Currently, the RT supports only plain inference with user provided masked inputs like:

<QED>0.123|[C][MASK][C][O]

which is prohibitive, especially since we use SELFIES. Til now, scaffold decoration was impossible with selfies since smiles-selfies conversion mappings could not be retrieveid. This is fixed in SELFIES >=2.1.0, for details see https://github.com/aspuru-guzik-group/selfies/pull/75.

Bumping selfies to >=2.1.0 requires an update of PaccMann models, so this issue is sequential to #38.

[jannisborn]

This is stalled until selfies 2.1.0 will be officially released.

[jannisborn]

Removing the stall since Selfies 2.1.0 has been released: https://github.com/aspuru-guzik-group/selfies/releases/tag/v2.1.0

[jannisborn]

Solved indirectly via sampling_wrapper feature. See PRs #77, #157, #163