test: fix two racy tests

[jonhoo]

record_nodrop and recorder_drop_staged deadlock permanently in CI. When it happens the test binary never exits, so the job runs until GitHub's timeout with no failure output. It shows up most often under cargo llvm-cov, but nothing about coverage instrumentation is special here; it just makes the whole suite slower, and the suite also runs mt_record_static/mt_record_dynamic, which keep 32 CPU-bound threads on a 2-4 vCPU runner. The reader thread is then far more likely to be preempted in the one window where this race is lost.

refresh() reads recorders under truth, bumps phase, and then blocks in recv() until it has received that many histograms. That recv() cannot fail, since the Sender lives in the Arc<Shared> the SyncHistogram itself holds, so the wait is unbounded. A live Recorder only sends when a subsequent write observes the new phase, or when it is dropped.

Both tests had the writer perform a fixed number of writes and then park on a Barrier that the reader only reaches after refresh() returns. If the writer got through all of its writes before the reader's phase.fetch_add, it never observed the phase shift, parked on the barrier, and could not write again -- while the reader blocked forever on a histogram nobody would send, and so never reached its own barrier.wait(). Neither side can make progress. Confirmed by forcing exactly that interleaving under gdb: breaking just before the recorders read is enough to hang both tests.

So the writers now record until the reader tells them its phase shift is through, which is the pattern clone_idle_recorder already uses. The handoff is explicit instead of depending on who gets scheduled first.

The cost is that record_nodrop loses its exact-count assertions: the number of writes is no longer fixed, so it can only check that the phase shift carried at least one sample across and that everything it carried was TEST_VALUE_LEVEL. What the test actually guards -- that refresh() is released by a write while the recorder is still alive -- is unchanged, and is now guaranteed rather than hoped for. recorder_drop_staged keeps its exact assertions, since the writer returns its own count.