Add percentiles to key metrics

Currently we report average and roughness for most metrics (like latencies). While these are generally useful, we should also report 99 and 99.9 percentiles. Otherwise it is very hard to observe things like tail latencies.

We probably don't need this everywhere - but it would be nice to get those at least for things like disk latencies, read latencies, and commit latencies.

No files, tests, or entry points are identified. Start by locating the metric aggregation and reporting code for disk, read, and commit latencies; done should include 99th and 99.9th percentile values for the selected metrics without requiring them everywhere.

Written by the indexing model from the issue text.