What API should be used for loading models?

[mattwthompson]

After #44 I expected there to be a fairly tidy step involving updating downstream consumption of this package's public methods, but I'm finding some surprises. (If your brain works better with examples, the concrete question I'm trying to get at is "where should I add cache-checking/model-downloading logic so that it's actually used?")

• openff-nagl never imports openff.nagl_models, so there isn't even code there to update to use dynamic fetching and caching
• openff-toolkit, for better or worse, uses validate_nagl_model_path as the chokepoint through which a model name (as a string that looks a lot like a PyTorch model) is magically transformed into a (full) path on disk that corresponds to a model which openff.nagl.GNNModel can gobble up
• some examples (but no released OpenFF packages) use list_available_nagl_models, but this is just to show what models are available. There isn't an obvious use of this method aside from just seeing what files are available. In any scientific work that I can think of, and even hastily-made examples, the user actually declares which model they wish to load.

A constraint we are arguably working under, for better or worse, is how the toolkit currently loads models. There are absolutely paths forward for changing that in favor of better plumbing, but it does imply that we could have validate_nagl_model_path be the point through which all logic and/or point users toward. (We might want to rename this function, but that can be done in a backwards-compatible manner with an alias.) I have a little prototype that explores this idea in a manner that might not require any code changes in other packages (#58)