Improve test runner parallelism

[lightmare]

This is an attempt to more evenly distribute testing work among child processes.

The first thing I did was get rid of the longRunningWorker. These tests take ages to complete, so I don't understand what's the point of running them all in one process.

Then I limited the number of tasks assigned to each worker at a time. Most of the tests were loaded and then buffered in the supervisor process anyway, once pipe buffers were full.

[devsnek]

so i gave this a few runs, and it seems to take an average of about seven seconds off my test run time. i also notice the ci time went up from 13 to 19 minutes. unless this is significantly faster for you i'd rather not merge it.

[lightmare]

When you run it without slow tests, it sometimes finishes a few seconds sooner — in particular when the original distribution gives some workers less work (time-wise) they finish early and you lose parallelism near the end of a run.

With slow tests, for me the original takes ~58 mins (3+1 workers), whereas this PR takes ~40 mins (3 workers).

I'll need to investigate what I messed up, because something's not right about the timings.
First, the CI seems to have stalled for twice as long near the end than it did previously.
Second, this PR didn't even print correct final counters (72323 + 3412 < 75737).
Third, summing up total CPU time spent gives me <100 min for the original, and ~120 min for this PR.

[lightmare]

Update: the issue with CI taking longer was probably caused by running 2 workers (i.e. 3 processes including the supervisor) on a VM with 2 processors. Trying to keep a short, but always nearly-full queue, introduced a lot of expensive context switches. I've fixed this by decreasing the queue length below which scheduleTests() loop in the supervisor wakes up.

The second issue (incorrect final counters) was a bug in this PR, fixed.

The third issue (more CPU time used) is in the engine or node itself, not in this PR. It happens on the main branch as well, which you can confirm by commenting this line in slowlist:

#built-ins/{encode,decode}URI?(Component)/**/*.js

and running:

time NUM_WORKERS=1 node --predictable test/test262/test262.js -- test/built-ins/encodeURI/ 

#######################
 engine262 Test Runner
 Detected 4 CPUs
 Not running on CI
#######################

[01:29|:   62|+   62|-    0|»    0] (0.00/s)

real	1m30.051s
user	1m29.123s
sys	0m1.185s

versus:

time NUM_WORKERS=3 node --predictable test/test262/test262.js -- test/built-ins/encodeURI/ 

#######################
 engine262 Test Runner
 Detected 4 CPUs
 Not running on CI
#######################

[00:50|:   62|+   62|-    0|»    0] (0.00/s)

real	0m51.201s
user	1m50.333s
sys	0m1.664s

When running the above I also measured time spent in some parts of the runner, and found that the 20-second increase in user time matches a 20-second increase in total process.hrtime spent in this call:

engine262/test/test262/test262.js

Line 308 in 8993f98

completion = realm.evaluateScript(test.contents, { specifier });