Fix error handling in MEGAN

[kmaziarz]

In search experiments, it can be observed that MEGAN sometimes produces odd molecules containing C+ atoms, which violate valence despite being accepted as valid by rdkit. These molecules fail sanitization through MEGAN's fix_explicit_hs utility, and so far our wrapper has been ignoring this and passing the molecule through to the model to propose reactions (likely pushing the problematic C+ atom to reactants). In this PR, the underlying model is not called on inputs which fail fix_explicit_hs, and rather an empty list of reactions is returned in this case. Relatedly, I also silenced some of the other internal warnings produced by the MEGAN model by applying a (now improved) suppress_outputs context manager.

[AustinT]

LGTM (although I only checked it superficially). It's nice to see a TODO comment in the code get dealt with!