Moore-Penrose -> least squares

[colinjcotter]

For manifold meshes, we use the Moore Penrose inverse of the gradient of the coordinate field. If the mesh element is anisotropic, this gradient can be ill-conditioned, and then evaluating A^TA is not a good idea or it's inverse is not a good idea.

Since we always immediately apply this operator to a vector, it would be maybe be better to just do a least squares solution of Ax=b, which does the same job, at runtime. Then we can use a stable algorithm like QR. This would require to keep the inverse symbolic, and insert a LAPACK SGELS or similar into the kernel code.