GeezerVison

This PR introduces a number of performance enhancements for our thin plate spline implementation. This came about from trying to represent and simplify mesh transformations originally derived from FEABAS.

Anecdotally, I was seeing ~30x single-thread speedups in 5000 control point runs in both fitting and simplifying the transformation. I am also seeing ~2x speedups in inversion (based on the new methods) when changing from the "gradient descent" fixed-point iteration method to newton iteration, mostly from much faster convergence.

Changes:

• Chunked vectorized deformation contribution and apply methods
• scale and translation normalization for fit method
• implement scipy.special.xlogy
• improved array management everywhere
• reduce matrix size in fit w/ x,y RHS columns
• cdist for sqeuclidean residuals
• consider symmetry in solve
• simplified gradient descent implementation
• optional newton iteration based inversion: more expensive, but faster-converging