Better default behavior for failure/success of MBAR solver.

[mrshirts]

This will need to pull in a few other issues.

• It could be that we need to make "robust" the default optimizer, with documentation about expected speed/peformance if people want better performance. (link relevant issues here)
• Right now, a lot of times failure to converge is not noticed, because the logger is not turned on, an failure to converge is not elevated to the top level. Probably a higher level exception should be thrown, since this is a huge failure mode.
• The results could also provide a .success attribute we can check? (like scipy's OptimizeResult) which can be checked.
• One problem that seems to be happening is that when one method fails, and it then goes to another method, it can use an unconverged, divergent previous guess, rather than starting over: hybr solver fails, then seeds adaptive solver with bad starting f_i #544