test(qwp): release test-owned native memory before the leak check fires (M7)

- BackgroundDrainerDurableAckRetryTest (the M7 finding): every
  ScriptedFactory allocated a StubWebSocketClient whose real super-ctor
  eagerly mallocs ~128 KB of native buffers (64K recv + 64K send + 256B
  control), never closed -- 18 tests, ~2.3 MB leaked per class run, and
  no assertMemoryLeak wrapping unlike its siblings. Stubs are now
  tracked in LIVE_STUBS and closed by a local assertMemoryLeak wrapper
  in a finally INSIDE the wrapped lambda -- LeakCheck fires at the end
  of the lambda, so an @After-only close would run too late. @after
  keeps an idempotent closeAllStubs() as a safety net for non-wrapped
  exits.

- QwpColumnBatchViewsTest: same anti-pattern -- 58 tests tracked their
  native allocations but freed them in @after, after each test's
  LeakCheck. Redirected to a local wrapper that frees inside the lambda.

- EngineCloseSlotLockReleaseTest: the sabotaged close() NPEs before
  reaching the watermark, and the manual cleanup captured ring+manager
  but forgot the AckWatermark's 16-byte mmap. Now captured and closed.

All three were invisible to the leak check because it tolerated growth
confined to a single memory tag; the next commit removes that tolerance.

Author: bluestreak01