Model checkpoints

[thopkins32]

Summary

Adds the ability save an optimizer checkpoint for the optimizers that implement the Checkpointable protocol.

AxOptimizer changes

• Saves checkpoint to a JSON file (configurable path)
• Can load a new AxOptimizer instance from a checkpoint

Agent changes

• Saves optimizer state to a JSON file (configurable path)
• Can load a new Agent instance from a checkpoint
• Removed a lot of unused internal state (and public views of it) to simplify loading from a checkpoint
• Rename evaluation argument to evaluation_function

optimize changes

• Can choose a checkpoint interval to save the optimizer state periodically

Other comments

Needed to simplify the internal state of the Agent in order to reconstruct it from a checkpoint. The alternative was trying to read private Ax attributes to reconstruct Blop's DOFs, Objectives, constraints, etc. which is not trivial.

We had to use private member access once to get the parameter names of the loaded Ax experiment. Apparently there is no good alternative to this.