Release 3.0

• Separates out the self-play, arena, and train into discrete processes.
• Improved Arimaa implementation with latest composite move engine.
• Update training loop to tensorflow 2.0 with custom gradient tape.

Release 2.0

• Adds Policy Softmax Temperature to act as a regularization force on policy.
• Converts Policy output to be direct logits of a convolutional layer in place of a fully connected layer.
• Adds additional ancillary heads like moves left.