std/atomics: lock-free for any 1/2/4/8/16-byte type, not just Trivial

[elijahr]

Changes std/atomics to determine lock-free eligibility by size, supportsCopyMem, architecture, and memory management rather than by type.

Previously, only Trivial types (SomeNumber | bool | enum | ptr | pointer) could use lock-free atomics. Any other type (small objects, char, etc) would fall back to spin-lock even though it could fit in a register.

Now, any type with sizeof(T) of 1, 2, 4, 8, or 16 bytes (with architecture support) and no managed memory (under ARC/ORC) uses lock-free atomics.

Dependencies

This PR depends on #25374 which fixes a compiler bug that blocks using when isLockFree(T) in the Atomic[T] body.

Changes

• Replace Trivial type constraint with public isLockFree(T) template
• Add hasLockFree8 constant for 8-byte atomic support
• Add hasLockFree16 constant for 16-byte atomic support
• Add 128-bit atomic type (Int128) for 16-byte lock-free operations
• Add -d:nimEnforceLockFreeAtomics to error instead of silently falling back to spinlock
• Add -d:nimNoLockFree16 to disable 16-byte lock-free for old x86-64 CPUs
• Document lock-free eligibility rules

Architecture Support

8-byte atomics (hasLockFree8)

Architecture	Supported	Instruction
All 64-bit	✅	Native
x86-32	✅	CMPXCHG8B
ARM32	✅	LDREXD
MIPS32	❌	-
PowerPC32	❌	-
SPARC32	❌	-
RISC-V 32	❌	-

16-byte atomics (hasLockFree16)

Architecture	Supported	Instruction
x86-64	✅	CMPXCHG16B
ARM64	✅	LDXP/STXP
Other 64-bit	❌	-
All 32-bit	❌	-

Note: Use -d:nimNoLockFree16 if targeting very old x86-64 CPUs (pre-2005) that lack CMPXCHG16B.

Lock-free eligibility by type and memory manager

Type	refc	markAndSweep	none	arc	orc	atomicArc
int, bool, char, ptr	✅	✅	✅	✅	✅	✅
enum, distinct	✅	✅	✅	✅	✅	✅
Small object (1-4 bytes)	✅	✅	✅	✅	✅	✅
8-byte object*	✅*	✅*	✅*	✅*	✅*	✅*
16-byte object**	✅**	✅**	✅**	✅**	✅**	✅**
ref, string, seq	✅	✅	✅	❌	❌	❌
Big object (>16 bytes)	❌	❌	❌	❌	❌	❌

* Requires hasLockFree8
** Requires hasLockFree16

[elijahr]

Converting to draft - still working on the API design.

• Flipping the default so non-lock-free atomics error at compile time (use -d:nimAllowAtomicSpinlock to opt into spinlock fallback)
• Also expanding documentation on safe usage with managed types and different memory managers