Prediction grid on the fly

[kyle-messier]

Potential "on-the-fly" alternative to prediction grid

We need to deliver predictions at fine-scale locations, but we don't need them all at once. For example, 1-km variability is not visible in any map over large regions like the entire US.

Permanently available predictions require 10+ billion covariate calculations and predictions, each.

On-the-fly trades of storage requirements with the need to orchestrate calculations where/when needed. Some things we'd have to consider in doing on-the-fly predictions:

1. Covariate calculations prediction locations
2. Base-learner predictions and base-learner auto-regressive predictions required
3. Could potentially require frequent calls to the {targets} pipeline
4. Scope of current {targets} pipeline may be different