Update constraint assembly

[ddement]

This is a series of commits which make three major changes to the way that constraints are assembled into the global system. The result is a significant improvement in performance, as highlighted in issue #363. The major changes are highlighted in the three main commits:

1. The transpose of the constraints matrix is now done at constraint variable calculation time rather than trying to perform this operation on the sparse matrix itself.
2. The components of the residual arising from multiplying lambda with the transposed constraint matrix are now computed at constraint variable calculation time rather than using a spgmv operation. Similarly, the non-transposed constraint matrix contributions are now multiplied by the tangent matrix at variable calculation time rather than using a spgemm algorithm.
3. All of the work done in CalculateConstraintResidualGradient is now done in local stack memory, just copying to/from the global heap memory at the beginning and end.

The PR looks massive because Git is interpreting some of the structural changes as complete rewrites of those files. To understand the meaningful changes, look at calculate_constraint_residual_gradient.hpp, assemble_constraints_matrix.hpp, and assemble_constraints_residual.hpp.