Interface for exploration policy

What would be a good interface for specifying the exploration policy?

It is implemented differently here and in `DeepQLearning.jl`.

- What is implemented here: Just allows a limited set of possible policy e.g. `EpsGreedyPolicy` and uses the internal of that policy to access the Q value. I think it is pretty bad: `EpsGreedyPolicy` should be agnostic to the type of policy for the greedy part (right now it assumes a tabular policy I think), if we improve `EpsGreedyPolicy` then the code here will break.
- In `DeepQLearning.jl`, the user must pass in a function `f` and `f(policy, env, obs, global_step, rng)` will be called to return the action. I took inspiration from MCTS.jl for this. However it is not super convenient to define decaying epsilon schedule with this approach.
- A suggestion is to use a function `action(::ExplorationPolicy, current_policy, env, obs, rng)`. Dispatching on the type of `ExplorationPolicy` and having users implement their own type seems more julian than passing a function. The method `action` is not super consistent with the rest of the POMDPs.jl interface since it takes the current policy and the environment as input.

Any thoughts?