Investigate why CategoricalClassification needs so many more MonteCarloPosteriorCollection samples than other classifiers

[rg936672]

What's the issue?

Identified in #379. CategoricalClassification controllers will fail on classify_fuzzy_points with a "covariance matrix is not positive definite" error if fewer than ~90 samples are used. This isn't an issue to end users (since by default 100 samples are used in all cases), but every other decorator works fine with as few as 20 samples, and being able to cut this down similarly would improve the runtime of test_decorator_combinations.

[tp832944]

Causes the tests to run slowly. Pretty low priority.