fix bwos queue bug

[viktorika]

1.Running TSAN on my M3 computer reports a data race. I believe there are some memory ordering issues with thief_block_, tail_, and steal_tail_.
2.There is an extremely small chance that elements within a certain block might be skipped and never executed. The reason is that the steal function uses steal_tail to determine whether it is complete or empty, but steal_tail == block_size might indicate that the current block has already been taken over.

std::uint64_t spos = steal_head_.load(stdexec::__std::memory_order_acquire);

Using steal_head for checking can avoid this issue.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ericniebler]

/ok to test f964a97

[ericniebler]

@maikel could you take a look please?

[viktorika]

Please check this issue. I have described the problem in more detail.#1755

[maikel]

@viktorika Does #1967 fix your issue?

[viktorika]

@viktorika Does #1967 fix your issue? #1967 解决你的问题了吗？

@maikel
I believe this implementation is correct, but does it create any efficiency issues if the thief always needs to start from start_block_ to find a block to steal?

[maikel]

That difference would need to be measured. I don't think that its a big issue, but intuition is often wrong here.

More imprtantly is, that we have correct version first. Dropping tasks was a big nogo

[ericniebler]

what is the status of this pr? is a fix still needed here? attn: @viktorika@maikel

[maikel]

I believe my other changes a month ago have fixed this issue here