Should bootstrapping be implemented at the alchemlyb level, or pymbar level? For MBAR, it would be better at the pymbar level, since it can be easily encapsulated (user doesn't have to worry about it), and one can request either uncertainty estimate.

For BAR over several states, then the bootstrapping needs to be done at the level ABOVE the BAR call, since we need to bootstrap all of the data simultaneously before feeding it into BAR. Same for EXP applied to a string of states.

Thoughts?