Faster  BlockRandomAccessSparseMatrix::SymmetricRightMultiply.

Trade a small amount of memory to improve the cache coherency of
the SymmetricRightMultiply operation.

The resulting code leads to a 10-20% speedup in the linear solver
end to end.

Change-Id: I8ab2fe152099e849b211b5b19e4ef9f03d8e7f1c