Dev/add caching for comp ref inference

[dmjoy]

(NOT INTENDED TO BE MERGED IN CURRENT STATE)

Add some form of caching to both the comparative regression KDMA estimation, and action parameter filling (with the intention of running through many many alignment targets without needing to re-run inference). The way this is set up we have to specify the dependencies for a cache entry manually (but that gives us a bit more power/flexibility; and other common ways of caching in Python are either only in memory (the stdlib approach), or require pure functions instead of methods). Let me know if I've missed any "dependencies" here for each function.

I'm on the fence about whether to clean this up and try to merge it in to main, but I'm concerned that we might not be covering every case well here and use the cache when we didn't intend to, etc.