Determine general API structure for BMC for general applications

[kylegodbey]

We can use this issue to detail how we imagine interacting with our BMC framework in a general purpose way. Our existing implementation is very much hardcoded for mass tables, but I think it isn't too hard to come up with a general structure. This is sometimes called spec-driven development (nice writeup: https://medium.com/@ratherabstract/sdd-spec-driven-development-3556dacca165), and it essentially means we craft the specification and then use that as a guiding light for development.

To kick it off, we should break the implementation we have now into its core components:

1. Set of models
2. Valid on a given domain X
   2a. X can be a vector
3. With a prediction Y3a. Y(X) can, in principle, be a vector itself (masses, radii, etc)
4. Optional orthogonalization step
5. Training, validation, and optional testing sets defined on a subset of X
   5a. It's also possible that we don't have a full Y(X) vector across each of these sets
6. A training method that determines model weights given a training set
7. A prediction method that, given a valid X vector, produces Y with the learned model weights

Am I missing something? Tagging @syiman as well for his insights

[kylegodbey]

This was started in #3 . It's still missing implementation. @turtuledunnad, if you could make a sample.py file that is a minimal working example of the bmc framework, we may be able to move quickly on this. Essentially you can convert your jupyter notebook to a python file (using nbconvert, for instance) and upload it here. It will be a good sample for us to pull from.