Add an RL model and associated files

[rahuls-cerebras]

A modelzoo RL model wrapper which contains a configurable policy_model
Add a dataloader to read npz files.
Add a hack to store and load checkpoints only for the policy model and not for the wrapper RL model.
Add a policy loss function to calculate loss.
Add a rl model registry

TODOS

• Think about how to communicate the MSL.
• It was not required in the existing flows because data preprocessing logic adds padding tokens to ensure the tensors are of size MSL.
• But in the case of verl, although the MSL is 128K, the tensors floating around are 4K.
• Think about how to have a permanent solution for the saving and loading of checkpoints