migrate to Chase-Lev deque

[tzcnt]

There are 2 versions of the implementation, chosen based on the target platform.

• 32 bit: the standard Chase-Lev implementation
• 64 bit: packs both indexes into a single word so that we can accurately count the number of elements available in the queue with 1 operation. try_pop() has a fast path that completes in 1 operation when we can confirm there was more than 1 element in the queue.

Removed the old ExplicitProducer-based work stealing setup which was providing the equivalent functionality. The moodycamel queue contains only the ImplicitProducer-related code, and is still used to accept work from external threads.

---

Breaking changes: TMC_WORK_ITEM=FUNC is no longer supported, since the Chase-Lev queue requires the work item type to be trivially copyable and trivially destructible. Users of this configuration should migrate to TMC_WORK_ITEM=FUNCORO instead.

[solbjorn]

Tested this PR on my engine:

• Windows 11 x86_64 26220.8544
• Clang/LLVM 22.1.7
• libc++ and llvm-libc 22.1.7 (unstable ABI v2, hardening, hardened iterators)
• -O3 -flto -fwhole-program-vtables -std=c++26 (+ no MSVC compat etc.)
• TMC configuration:

#define TMC_NODISCARD_AWAIT
#define TMC_PRIORITY_COUNT 3
#define TMC_STANDALONE_COMPILATION
#define TMC_TRIVIAL_TASK
#define TMC_USE_HWLOC // hwloc master branch
// #define TMC_WORK_ITEM=CORO -- default

• tmc::cpu_executor() + standalone tmc::ex_cpu_st

No compilation issues or new warnings, no runtime issues, no perf regressions.
exe size decreased by 12-15 Kb.

The diffstat is also nice.
Rocking as always \m/

[tzcnt]

Thanks. Performance in all benchmarks has been equivalent to the old queue, but it should result in less cache miss / TLB pressure overall, and the code is shorter / easier to understand.

I'm going to delay merging this to main because it's technically a breaking change (requires the removal of TMC_WORK_ITEM=FUNC), and I'm considering releasing it along with some other breaking changes.

[tzcnt]

I changed my mind; I think this can go with v1.6:

• I doubt anyone's using TMC_WORK_ITEM=FUNC since the performance was poor anyway
• It's a relatively small breaking change that is easy to fix (swap to using TMC_WORK_ITEM=FUNCORO instead)
• It can be detected at compile time with a static_assert