modrax

Improve evaluation

Added multilayer RNNs and proper train_forward

Added support for recurrent network cached forward + evaluation against uniform of recurrent network

Add gumbel alphazero algorithm and example

Added scaffolding for custome val function