In #93838 we started using a seeded rand for use in the load based +
splitter's reservoir sampling algorithm.

rand.Soure heap allocates on init approximately 4kb. When intialized
per-replica, this is problematic as nodes scale to large replica counts.

The splitter should instead be initialized with the global rand, to reduce
memory overhead per-replica.

xref #94737

Jira issue: CRDB-23121