Speed up large group by (top k) queries

[mtanski]

Doing groupby string column (domains in reverse order) is one of our worse
performingqueries we have. Mostly because we're grouping by 10s and 100s
thousands of unique values. I've been trying to make this perform better for
a long time. We got a nice but small speedup (5%) before by using CityHash64.

After spending more time with perf, it looked like of the if condition branches
was always taking a lot of time. For some reason it kept getting miss
predicted. It was a simple boolean condition.

On very large groupby (100,000 elements) this improves execution time by 5% to
10%. No observable regression to small inputs once the hash size check was
introduced.

[mtanski]

It wouldn't hurt to use CityHash64 too. It's faster then the included Murmurhash64 about a net %5 shaved of our large group by queries with no other regressions. Works well for both large and small strings.

Also tried FarmHash but it performed noticeably worse for short strings (like domains) then CityHash.

[mtanski]

Actually, I'm seeing the small integer group by case (~ 250 values) get faster on a large number of runs by like 1.2%.

[mtanski]

Despite all the other work (CityHash), LRU buckets, moving the HASH_COMP_ONLY branch out (template arg). The end of the linked list check is the largest time sink. Not sure what we can do about this branch.