Faster kernels for assembling mass matrix

Solves the following issue(s):

Partially sovle #253 (closed)

Core changes:

Improve kernel_3d_mat by using smaller array to store part or single dimension of multidimensional array.

The kernel is now approximatively 33% faster (went from 0.2535s/exec to 0.1672s(exec) in my variational tests).

Model-specific changes:

Fix a small mistake in the variational propagator where the error was computed after the update and hence always being close to zero instead of the real error.

Documentation changes:

None