Convert ComplexMat to use DynMem class

This is the first step to unification of CPU and GPU implementations.

It also fixes a incorrect dimensions of ComplexMats in
Gaussian_Correlation, which was unnoticed, because the matrices were
reallocated on the fly. Now, we use assertions in DynMem, which
prevents this. Ideally, we don't want any allocation during ComplexMat
computations, but we're not there yet.