Exact score based algorithm optimizations

[mwien]

Some fairly free optimizations. I benchmarked briefly for graphs with 16 vertices.

• avoiding the unnecessary memoizations by calling local_score_ directly: reduces execution time from ~2.5s to ~1.2s (so roughly 1s faster)
• removing @view which I think can hurt because the submatrix is not contiguous: reduces execution time from ~1.2s to ~1.1s

We should definitely fix the first point (that was just an oversight that these calls are memoized)...

Second point we could benchmark GES on larger graphs first

UPDATE: the @view thing doesn't make a big difference for GES, so I have no big preference whether we keep it