Action masking feature (legal actions)

POMDPs.jl [supports state-dependent action spaces](https://juliapomdp.github.io/POMDPs.jl/stable/def_pomdp/#state-dep-action)

However, DeepQLearning.jl is always picking the full action space.
That's because the `solve` enumerates the actions once [here](https://github.com/JuliaPOMDP/DeepQLearning.jl/blob/e62ac7626c7a2f4b2b981208a3bb557e71f7d327/src/solver.jl#L41C5-L41C15), hands them into the policy, which are broadly used there after.

Do you think of a way to have action masking with the current implementation ?

Start in src/solver.jl at the solve code around line 41, where actions are enumerated once, then trace how the policy uses that action list. Review the POMDPs.jl state-dependent action-space behavior as context. Done means DeepQLearning.jl can select only legal actions for each state without breaking the current policy flow.

Written by the indexing model from the issue text.