timers.wakeTime, which is called concurrently by P's trying to decide
how long they should sleep, can return inaccurate values while
timers.adjust is running. (Before the refactoring, this was still true
but the code did not have good names and was spread across more
files, making the race harder to see.)
The runtime thread sleeping code is complex enough that I am not
confident that the inaccuracy can cause delayed timer wakeups,
but I am also not confident that it can't, nor that it won't in the future.
There are two parts to the fix:
1. A simple logic change in timers.adjust.
2. The introduction of t.maybeAdd to avoid having a t that is
marked as belonging to a specific timers ts but not present
in ts.heap. That was okay before when everything was racy
but needs to be eliminated to make timers.adjust fully consistent.
The cost of the change is an extra CAS-lock operation on a timer add
(close to free since the CAS-lock was just unlocked) and a change
in the static lock ranking to allow malloc while holding a timer lock.