Optimize model writing and loading during prediction

[blythed]

We should have separate processes writing model outputs, while other processes are loading data and predicting.
This is particularly obvious when the inputs and outputs are large (order of magnitude of work in I/O comparable to prediction).