proposed fix for dynamics back-prop

[bilkitty]

reward preds are not detached so as to yield policy_loss.require_grad = True (see diff in commit 7677f5c)

[jurgisp]

It is important that the reward predictions are detached.

[bilkitty]

Thanks! That was a bad guess.

Going back to questioning the assert, I've updated the branch to enforce it only when using reinforce.