feat(sync): added atomic<T> and atomic_ref<T> wrapper primitives over raw atomic builtins

[FlareCoding]

Summary

• Atomic access was spelled as raw __atomic_* builtins with a memory order argument repeated at every call site, making orderings easy to get wrong and overall looked messy as hell.
• sync::atomic<T> (owned, non-copyable) and sync::atomic_ref<T> (borrowed view of plain storage) encode the ordering in the method name and reject non-lock-free types at compile time.

---

Note

Medium Risk
Changes memory-order spelling and refcount/futex/spinlock synchronization paths; semantics are intended to match prior builtins but mistakes here affect core locking and object lifetime.

Overview
Introduces sync::atomic<T> and sync::atomic_ref<T> as typed wrappers over GCC __atomic_* builtins, with memory order encoded in method names (load_acquire, fetch_sub_release, etc.) and compile-time lock-free requirements for 1/2/4/8-byte integers, enums, and pointers. atomic_fence_acquire / atomic_fence_release replace ad-hoc thread fences.

ref_counted stores the refcount in sync::atomic<uint32_t> instead of a plain uint32_t with raw builtins; try_add_ref now treats REFCOUNT_POISON like zero (returns false), and the last-ref path uses sync::atomic_fence_acquire before poisoning.

Ticket spinlocks in spinlock.h use atomic_ref<uint16_t> on next_ticket and now_serving with the same relaxed/acquire/release pattern as before.

futex_wait re-reads the user futex word under the bucket lock via atomic_ref<uint32_t>::load_relaxed() instead of memcpy, keeping the atomic check-and-enqueue behavior while spelling the read as an atomic load.

^{Reviewed by Cursor Bugbot for commit 9a5eb48. Bugbot is set up for automated code reviews on this repo. Configure here.}