Replica networks cleanup - #263

Passes Travis! This is now ready for review and merge.

I still find the specific question asked in #262 a little problematic (see my most recent #262 (comment)), however, I've completed everything I set out to do in this PR.

Looks good.

Still thinking that we might make is_ensemble_change_mover computable and not having to specify it directly.

How would we make it computable? And does that make sense?

Right now, you only have to do anything if you're creating a SampleGeneratingMover that also involves changing ensembles. Any combination movers already compute it from their submovers.

And consider this minus-like sequential move:

1. Take trajectory from innermost interface, extend it forward until it satisfies the minus ensemble
2. Select the new subtrajectory that satisfies the innermost interface.
3. Replace the old innermost interface trajectory with that new one.

For this, is_ensemble_change_mover should be False. Currently, the user would have to set that manually, but my approach allows the user to set that manually. (This is why it is extremely important that the default be None, not False). A computational approach would need to figure that out as well.

This approach also allows a user to exclude certain ensemble-change movers from analysis; for example, maybe you want to look at only the normal replica exchange, and none of the minus moves. You can set the flag to "False" on the minus moves before doing your analysis.

How would we make it computable?

I don't know, maybe that is not possible :) Just thought it would be cool, if we could avoid setting it and derive it from some properties.

And does that make sense?

Why not, if it works without fiddling? Then we could remove the code and a potential user that wants to write a new Mover does not need to worry about it.

This approach also allows a user to exclude certain ensemble-change movers from analysis; for example, maybe you want to look at only the normal replica exchange, and none of the minus moves. You can set the flag to "False" on the minus moves before doing your analysis.

I think this is the important point. it is for analysis to manually set, what you are interested in. A computed approach does not have this flexibility. Although it also changes the meaning a little.

So maybe move that idea to distant future :)