Update `R.nzval` instead of `R[i,j]=...`

Currently, we use R[indices[i], j] = submatrix_result[i, idx] to write the submatrix result into the result matrix R. This is suboptimal because we know that only R.nzval will actually change eventually (the sparsity pattern is preserved) and, more importantly, because setindex! doesn't seem to be thread-safe for SparseMatrixCSC. (To temporarily circumvent this issue, we currently allocate R as a dense matrix only to make it sparse before returning it...)

At least in the non-serial case, we want to update (the respective part of) R.nzval directly. This should be

• a bit more efficient (how much?)
• thread-safe (if different threads only operate on non-overlapping regions)

It is also what Michael did it his original work (see values_inv).