stats: reduce overhead of distinct estimation

[yuzefovich]

We currently use hyperloglog library to estimate the number of distinct elements. I just collected a 50s cpu profile non-nil datum alloc (about the time it took for ANALYZE to complete) on a cluster that only had ANALYZE tpcc.customer running, and this distinct estimation is the most expensive part of the stats collection (this was on dbb0baa plus a revert of 2831511 and another commit to introduce a cluster setting for using nil or non-nil DatumAlloc in stats):

We should investigate whether it's possible to reduce this overhead. We recently upgraded the hyperloglog library, so there is no quick fix like that :/

There were some ideas floated around that we could avoid this expensive computation altogether for key columns if we were to scan the secondary indexes.

Related to #135988.

nil datum alloc

Jira issue: CRDB-47355