Repair lock-free queue correctness and reclamation

[darkautism]

Reworks the HP branch around a standard Michael-Scott queue publication protocol and fixes the correctness problems found during review.

Key changes:

• link new nodes with CAS on tail->next before advancing tail
• keep queue next and retired-list links separate
• use atomic pointer/int operations on GCC/Clang and MSVC
• protect dequeue head + successor with two hazard pointers
• prevent reclamation while producers may hold stale local tail pointers
• enforce hazard-slot ownership so manual and automatic dequeue cannot collide
• fix max_consume_thread=0 default behavior, allocation/error handling, cleanup, invalid tids and reserved payload values
• add status-returning dequeue APIs while keeping legacy wrappers
• add edge/MPMC/ABA tests plus ASan/UBSan/TSan and MSVC CI
• refresh Visual Studio project toolset and README

This intentionally keeps payload ownership with the caller; lfq_clean only reclaims queue nodes.