Enable warm-start model fitting with the Ax backend

[crasanders]

The old AEPsych backend supports warm-start fitting for models. That is, the previous model's hyperparameters can be used as the starting point instead of re-fitting from scratch. This makes the fitting process faster. We would like to enable this functionality in the Ax backend.

This functionality could be enabled by passing the previous model's state dict into AEPsychSurrogate, however some additional logic may be required on the Ax side to be able to extract the stat dict using model._get_state.

Users would specify how frequently the model should be warm-started vs fit from scratch using the refit_every setting in configs as is done in the old implementation.