leveldb/memdb: avoid repeated comparisons

[holiman]

When the memory database does a search through the skiplist, it descends one level whenver it either finds that the list is ended, or that the comparison fails (we've passed the item).
In the latter case, we descend a level, and a non-insignificant amount of times, the search on that level will also end on the same exact item -- since if the item was at the upper level, it's guaranteed to be present on the lower level too.

By adding one local int, to keep track of, basically, "an item we found on the upper level, which is known to be past the item we're searching for", we can avoid doing that comparison again.

Using this fix, plus a global comparison counter, I made these charts. It imports 30M items, each with a 32 byte key and a 1 byte value. At 100 points during this import, I dump out the average comparison ops performed in that range.

This is master

It executed it in 175.91 seconds, doing 1416736538 comparison calls in total.

This is this PR:

It executed it in 171.51 seconds, doing 1333783947 comparison calls in total.