Small fixes to work-stealing-queue

[vchuravy]

While working on #55542 I took a closer look at the Chase-Lev-Lê queue,
we implemented for the GC.

I noticed two things:

1. As noticed in https://wingolog.org/archives/2022/10/03/on-correct-and-efficient-work-stealing-for-weak-memory-models the Lê implementation is liable to wrap around. Following Whippet here and the original Chase paper to calculate the size of the buffer https://github.com/wingo/whippet/blob/a180be0cbd3a77003192e50d094657b1e0a7af8c/src/shared-worklist.h#L201
2. In the steal Lê uses a consume load, we mistakenly down-graded that to a relaxed load, when we should have upgraded it to a acquire load. It synchronizes with the release store in push.

[vtjnash]

In the steal Lê uses a consume load, we mistakenly down-graded that to a relaxed load,

In a GC language, that isn't necessary AFAIK

[gbaraldi]

Right but that code isn't in a GCed language it's in the C part of it.

[d-netto]

We have "a poor-man's GC" for purposes of managing the memory associated with the ws_array_t though.

I.e., whenever we need to expand a queue, we allocate a new ws_array_t, push the old one into the reclaim_set, and only after the end of the GC we go ahead and free all of the arrays in the reclaim_set.

I wonder if this makes it legal to use load_relaxed here.