Unify training logic inside agent

[muralx]

One more refactoring :)

• All training logic is now inside the trainable agent and can be tweak independent of the environment
• Added hook methods to the Trainable agent, and calls from standard arena to allow arena to be reused for training purposes
• Removed most logic from train.py except from model saving and rendering of rewards/losses (done with arena plugins: Thanks Cucu for those plug ins)
• Using 1, -1 for win/loss. In theory we can add something in the agent to scale numbers if needed
• I changed the step_rewards to user board_size^2, which seem to work better

Side note
After the refactoring I was able to train DExp(Diego Experimental) Agent on a 5x5b 0w which has a win ration of .95. In theory, you should be able to reproduce it due to (-i 42).

Training cmd: python3 deep_quoridor/src/train.py -N 5 -W 0 -d 0.9999 -e 1000 -i 42 -s
Play cmd: python3 deep_quoridor/src/play.py -N 5 -W 0 -t 100 -r matchresults -p simple dexppretrained

[muralx]

@alejandromarcu I removed the renderers as parameter of the play and replay method. Renderers are already part of the arena, so not sure you moved them as args of the method. Please let me know what I missed.

[alejandromarcu]

Very nice, great use of plug ins!

[alejandromarcu]

Nice, actually I was working on an agent and noticed that the agent didn't know if it's player 1 or 2, so it doesn't explicitly know where to go. I was adding it in the reset method but this works as well. I think in the future we may want to standarize all this to use Player.ONE from quoridor.py, but we can do that separately