Comm/compute overlap increases task lifetime under ugni

[ronawho]

I couldn't come up with a great title for this, but comm/compute overlap in
ugni (calling chpl_task_yield while waiting for comm) increases the lifetime of
tasks and limits the number of short-lived tasks we can create. This can easily
lead to OOM situations like in #11820. Creating short-lived tasks ends up
happening for things like reductions, which has caused us to lower our stack
size in most of our scalability tests.

The following program will OOM on 2 locales under ugni. It just creates 10,000
no-op tasks on locale 0, but we do a yield while decrementing the end count and
the task will go to the end of the queue, and new tasks will run. In the worst
case we have to cycle through all 10,000 tasks before we can retire the first
one (and with 8MB stacks that 80GB of task stacks.)

use Time;
config const trials = 10_000;

var t: Timer; t.start();
var a: atomic int;
coforall 1..trials do on Locales[numLocales-1] do { };

writeln((a.read(), t.elapsed()));

For short-lived (or at least minimal comm tasks) that don't have any user
induced yielding I think we should be able to create an arbitrary amount of
tasks without OOM.