Rewards are only set once, only need to accumulate them once.
No need to modify the accumulated rewards if they ae not set.

For unknown reasons, this is required for stable baselines
training to work correctly. It does not appear to impact other
usage as the agents are still looped over correctly after the
game ends - just in a different order.

This makes the env code less cluttered and better encapsulates
the board behavior. It also expands the checks for a valid move.

These never change. There is no reason to recalculate them constantly.

removes the duplicate calls to check winner

This keeps the env from needing to access the internals
of the board to get the moves available.

This causes problems with test cases that have invalid
board configurations to test win lines.

This now always switches agents in each step. The previous change was a
bug that "fixed" the behaviour due to a bug in the SB3 tutorial.

The agent should be swapped every step as is done now.

The change to v4 was due to the agent handling at the end of an episode

The other changes to the env don't change the behaviour of the env, so
it is left at v3.