[BUG, ENH] Make sure classifier is new everytime and Add optional caching to classifier-based CI tests

[adam2392]

Problem

The classifier-based CI tests are nice because they are non-parametric generally, but slow because for each test, you need to re-fit a classifier. For modular components, we should optionally cache predictions.

For example:

1. CCMI: we can cache MI estimates of I(X;Y), I(X;Y,Z), such that we can re-use them when needed
2. CCIT: store the metric (and optionally biased metric) values for combinations of (X, Y, Z). The ordering of (X and Y) does not matter here.

Another issue is that we should ensure that the classification model is refit from scratch each time ci_estimator.test(...) is called. Downstream classification models in scikit-learn might support warm-starts, and this may be the subject

Proposed solution

Create a nested dictionary

For CCMI:

• X variable
• Y variable
• Z variable (can be None to indicate no conditioning)

For CCIT:

• X variable
• Y variable