`det()` implementation

This uses the LU factorisation provided by MAGMA to implement the determinant in the same way that Armadillo does (roughly). There are no specializations for tiny 2x2 or 3x3 matrices, though (I didn't think they'd give much speedup on the GPU. If your matrix is only 2x2 or 3x3 the GPU is not the right place for it). I had to implement some new kernels to compute products of vectors or products of diagonals, in order to compute the determinant of the factorised U and P matrices.

This MR depends on !80 (merged) and so the diff won't make sense until that is merged.