ENH: Performance improvements for generating candidate models

[bocklund]

This set of changes improves the performance of parameter selection with two primary changes:

1. When we build candidate models (renamed build_feature_sets to build_candidate_models) we take all combinations of the product of composition-independent features with interaction features. The implication of this is that some models that have a lot of features, for example heat capacity temperature features with four binary interaction features, can get very expensive to generate candidate models because the current implementation has geometric complexity with respect to the temperature and interaction features (as documented). Here we make an optimization for cases when the general implementation will generate more than complex_algorithm_candidate_limit (default =1000) candidate models, where the simplified version will have the same number of composition-independent features for all interaction features. Instead of geometric complexity $N(1-N^M)/(1-N)$, the simplified version has complexity $NM$, where $N$ and $M$ are the number of composition-independent features and interaction features, respectively.

2. A profiling-guided optimization in espei.paramselect._build_feature_matrix. The feature matrix is a concrete matrix of reals (rows: observations, columns: feature coefficients). We use a symengine vector (ImmutableDenseMatrix) to fill the feature matrix row-wise, moving an inner loop to fast SymEngine rather than slow Python. Roughly 3x speedup of this function after this change.

[bocklund]

Note that the performance issues this resolved are mostly due to a combination of the number of candidate models and the amount of data.

Generating the candidate models has an up front cost, but the result is cached so it's not overly expensive. The main contributor is that with many candidate models and data, most of the time is eventually spent in _build_feature_matrix. Since it's not entirely clear that the existing approach, while more general, actually generates better models, i.e. ones that are actually selected by the mAICc. In my experience, I haven't often seen models generated that have different features selected across different interaction parameter orders for the same interaction.