Initialize new run w/replay buffer of old run

[jonbinney]

Instead of just the model from an old run. There are two reasons for this:

1. If we initialize with the old model but an empty replay buffer, then we spend a bunch of time overfitting to the first few self play games the new run creates.
2. When we change the structure of the model or the features, it isn't possible to directly use the model weights from a previous run. In this case, we now have the option to start with a replay buffer containing relatively "Smart" runs to train the model with the new structure.

I'll leave this up for a couple days before merging in case @adamantivm @alejandromarcu want to add any thoughts