Exposed args to ErgoExo, provided _mflow_func abstraction

[jpbrodrick89]

ErgoExo now takes args as an additional argument to __call__ and value_and_grad, enabling another way to "train" opposed to modules.

Abstracted __call__ by adding _mlflow_func abstraction which could work for value_and_grad if we break backwards compatibility by changing return type slightly (e.g. packaging grad, val in a tuple), happy to discuss.