Experiment with WorkStealing

[vchuravy]

Basically a continuation from #43366.

It shows similar promise for pfib, but I was actually motivated
by the benchmark in #50202 (comment)
where we did a lot of work trying to figure out if there are no more tasks to run.

One issue I am aware off is that the underlying data-structure is not GC friendly.
Fundamentally we will only copy over data in the buffer, we are gurantueed to only
read valid data, but on the unused pieces of the buffer we have task-corpses laying
around.

A second issue is that the current implementation only allows for one thread to push! and popfirst!,
while all other threads are only allowed to steal!.

It looks like Go uses a double-ended queue.

[vchuravy]

For my own record (with 4 threads).

WSQueue improved pfib from ~2.1s to ~1.4
ILLS decreased pfib from ~2.1s to 5.9s

I am not sure how to make WSQueue fully correct so I will probably investigate a lock-free double ended queue.

[gbaraldi]

So this triggers an assertion when building with more than one thread set because

    if (jl_base_module == NULL) {
        // nthreads > 1 requires code in Base
        jl_atomic_store_relaxed(&jl_n_threads, 1);
        jl_n_markthreads = 0;
        jl_n_sweepthreads = 0;
        jl_n_gcthreads = 0;
    }

doesn't set the threadpool stuff, though I guess maybe it should @vtjnash?