Support for state-dependent action spaces

[Steven-Adriaensen]

Currently, there is no way to specify that certain actions are only possible in certain states.
Clearly one could just give a large penalty, but providing a way to model A as a function of current state s would definitely simplify modeling and allow solvers to exploit this feature.

One particular use-case would be controlling multiple (possibly conditional) parameters. In many cases it makes sense to model this such that at every step, only the value of a single parameter is changed, i.e. the actions would be the alternative values for the parameter to be set.