Need to run performance tests on $percentile in window functions with large windows. 

Some considerations for improvements could be 

1. Computing all requested percentiles at once.
2. Iterating in reverse for p > 0.5.
3. Change how we store _values (potentially use absl::btree_multiset and not std::multiset). This could decrease memory usage.

Run $percentile on window functions on queries and measure the latency.