sqlsmith: use actual data for generating predicates

Currently, whenever `sqlsmith` generates predicates, it simply uses randomly-generated datums. However, this can make those predicates less interesting if the random constants never match what we have in the database against which the stmts will execute. We should figure out a way to somehow use actual data from the tables to use as samples for random datums.

Jira issue: CRDB-25219

Start by locating the sqlsmith predicate-generation entry points and the code that executes generated statements against database tables. Determine how actual table values could be sampled for predicate datums; done means generated predicates can use matching data from the target tables and produce more interesting executions.

Written by the indexing model from the issue text.