[Wheel] Fix promotion-on-hit e2e flakiness from cross-class segment leaks (#3466)

Each test class in test_promotion_on_hit.py mounts a 32MB segment in
setUpClass but never unmounts it. Later classes therefore run against an
inflated master mem_total_capacity (64MB instead of 32MB), which pushes
used_ratio right onto the 0.90 eviction high-watermark boundary and
makes the promotion watermark gate timing-dependent — the root cause of
the nightly test_promotion_after_repeated_hits flakiness.

Add tearDownClass calling store.close() to all three classes (including
the skipUnless-gated BenchPromotionLatency) so each class's segment is
unmounted when the class finishes.