Optimization for 64 bit loads and stores for 32 bit x86 on gcc

[Andrey Semashev]

The attached patch contains some fixes and optimizations for 32 bit x86 atomics support for gcc and its pretenders. In particular, the patch contains:

1. Fix for BOOST_ATOMIC_LLONG_LOCK_FREE macro value on 32 bit x86. The macro should have value 2 when cmpxchg8b instruction is enabled in compile time and 0 otherwise (as long as Boost.Atomic doesn't perform run time detection).
2. Fixed compile time cmpxchg8b presence check. The previous check only worked on compilers with sync* intrinsics support or on i686 class CPUs (but not later or i586).
3. Optimized 64 bit loads and stores in 32 bit x86. When the target/source memory is 8-byte aligned (which is the most common case) it is possible to avoid expensive "lock cmpxchg8b" and use regular loads and stores instead.
4. Added a few missing noexcept specifications.

[Andrey Semashev]

The patch with fixes and optimizations for 32 bit x86

[Andrey Semashev]

Note: I tried it locally on Linux with gcc and clang 32 and 64 bit. All tests passed.

[timblechmann]

(In [84288]) atomic: Optimization for 64 bit loads and stores for 32 bit x86 on gcc

fixes #8574