Stop the thread store’s clear paying almost a boot to reset statistics after a capture

Make borrowing the per-thread test store (ThreadConfinedStore, the store a test thread boots once and empties between cases) cost what its javadoc says it costs after a real capture, not almost what a fresh boot costs. Today a borrow that follows a catalog capture takes about 200 ms against a 272 ms boot on the same small schema, and nearly all of it is StoreStatistics.reset: the capture leaves about 670 columns with measured selectivity, and the reset issues one ALTER TABLE …​ ALTER COLUMN …​ SELECTIVITY per drifted column (truncate plus the leak census together cost about 6 ms). So a module that moved its per-case fixtures from booting to borrowing, as graphitron-lsp did, cut its boots from 207 to 54 and saw no wall-clock change. When this lands, the reset after a capture costs a small fraction of a boot and that move pays off.