Fix: TrustRegionSolver: use blocked matrices for fufem scalar basis assembler

valgrind complained here and it is right: in the internal loop of the assembler we loop over localMatrix.N() and M(). Since dune-functions uses flat matrices for the localMatrix the number of dof's is dim-times too large for a powerBasis. In order to spare a double implementation of the local assembler only for powerBasis, we can simply convert it in the lambda. valgrind is now happy here.