perf(app): keep a busy tracker's boards current for less CPU and memory

[GraemeF]

On a busy tracker the watcher drew every project's trees on every collection and told each board about every project, even when the probe had found nothing new. Now a collection draws nothing when no read finds its tracker has moved. It only vouches that each project it read is as current as the read. A collection that does draw tells the watcher only of projects that were read again, or whose beads' readiness changed because a blocker in another project moved.

The watcher's threads now allocate from one malloc arena. Its live heap was the same before and after the read change, but each thread's arena kept what that thread freed, so resident memory sat at about twice the live heap. The arena setting has no unit test, because its effect shows only in a measurement.

Measured on the reproduced workload in bench/boards: the watcher plus five boards, a 90 s warm-up and a 300 s window, with runs interleaved main, branch, main, branch.

	main	this branch
Watcher CPU time	45 to 47 s	20 to 22 s
Watcher peak resident	423 to 454 MB	289 to 304 MB
Boards' CPU time	102 to 120 s	68 to 74 s

The boards stay as current as before. They saw the same writes, with a median delay of about 1.5 to 2 s on both sides and a longest delay under 3 s. The longest time a board's freshness answer went stale fell from 14 to 22 s to under 3 s. CPU time spent in bd under the watcher is unchanged, because the watcher asks bd the same things.