[DomGarguilo]

Fixes #6587

Moves the tserver's unopened, opening, and online tablet collections into a new
TabletAssignmentTracker that owns them behind a single private lock. Callers use named
atomic transitions (addUnopenedIfNoOverlap, unopenedToOpening, openingToOnline,
awaitNotOpening, ...) instead of synchronizing on TabletServer fields. This resolves the
11 USO_UNSAFE_OBJECT_SYNCHRONIZATION findings from SpotBugs 4.10, removes the three-level
lock ordering, and makes the assignment state machine unit-testable. OnlineTablets is
folded into the tracker.

Most of the tracker is existing logic moved as-is. To see only the real changes:
git diff --color-moved=dimmed-zebra --color-moved-ws=allow-indentation-change upstream/2.1...unifiedLocking

Behavior changes

• One lock instead of three, so there is no lock ordering.
• Fixes the lock-order deadlock on the rolled-back split path, the same issue as Fixed possible deadlock issue during fixSplit #6588.
• loadTablet overlap error logging runs after the lock is released, using tablets captured
  under the lock. Its output is unchanged.

Testing: new TabletAssignmentTrackerTest; tserver unit tests; -Psunny ITs; SpotBugs
4.9.8 and 4.10.5 (the 11 findings are gone).

Co-authored-by: Claude Opus 5.5

[dlmarion]

I don't think it's guaranteed that the thread has actually started running here. I typically handle this by waiting on a CountDownLatch, and call countDown in Thread.run.

[DomGarguilo]

Yea I think you're right. A latch would still leave a gap though before wait() so I just opted for a polling loop to wait until the thread is WAITING. Also added a @Timeout to the test case to protect against that loop. I tried removing the notifyAll call temporarily and the test failed so I think things are correct here now.

[dlmarion]

Loogs good