While fixing WT-3030, we noticed that a lot of searches saw ref->pindex_hint not match in the parent index being searched. The existing loops (from the hint to the end of the index, then from the beginning of the index to the hint) is inefficient in cases were the hint was slightly too large.

Also note that in workloads where an internal page is splitting frequently, there can be concurrent searches in different versions of the index. This contributes both to hints being incorrect and churn when readers overwrite the hint set during splits.