Memory requirement for dense systems

[flx-wlf]

In v0.99, dense systems are assembled via
A = T'( sum(A*(i,j))T)
with A*(i,j) being the basis-interaction of elements i and j, matrix and T the transformation matrix.
This means A* is assembled fully and then reduced by T, thus creating exhaustive memory overhead.
Although it is harder to parallelize properly, we should try to assemble it via
A = sum(T'A*(i,j)T).