Formula performance improvements

[nsmith-]

The current formula implementation walks an AST on evaluation. The AST is not optimized beforehand.
We could improve this in a few ways:

• Optimize the AST
• Convert the AST into opcodes and run in a VM (is making our own stack better than AST walk? WIP attempt at master...formula_perf)
• Find a way to statically link against llvm and compile to real machine code

A simpler solution might be to check if the formula is some regular pattern and convert it to a more specialized node. For example, we could create dedicated Horner-form polynomial evaluation nodes from formulas that are just multivariate polynomials. There is some literature on multivariate Horner forms e.g. https://arxiv.org/abs/1207.7079 and an existing python package: multivar_horner.