Remove redundant snake computations

[ninioArtillero]

Current implementation (of dstep) deviates from the paper in that it extends each candidate along the available snake before comparing matching candidates.

This PR contains a refactoring that computes the extension along the snakes after matching candidates have been merged, thus voiding the computation of a significant amount of redundant snakes. Indeed, given that addsnake preserves a node's diagonal (as proved in #29), both candidates are know to be on the same diagonal before extending them, so we necessarily get the same result.

The change gives an improvent of ~40% reduction of cabal bench time (almost twice as fast), and cabal test passes.