Support for state-dependent action spaces

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.

The issue names no files, tests, or entry points. Start by locating the environment and action-space APIs, then trace how solvers consume available actions; done should allow valid actions to depend on the current state and support conditional parameter changes without penalty workarounds.

Written by the indexing model from the issue text.