Connect the individual prediction gradients into the Combination workflow

[kaminow]

Currently the Combination implementations only return the final prediction, so if we want to calculate a loss against the individual pose predictions, we get an error since the computation graph for the individual predictions are empty. Want to connect in the individual pose prediction gradients to allow using a loss against these pose predictions (eg cross entropy loss for selecting the best pose).

[kaminow]

this might just be as simple as returning the individual pose predictions in the Combination.forward methods and then returning their gradients in the Combination.backward methods

[kaminow]

@jchodera @hmacdope I think I'm pretty happy with the level of commenting/documentation now, but please take a look when you get a chance and let me know what you think!