Fix CLockFreeGuard, introduce AtomicMutex

[prasannavl]

/kind fix

• CLockFreeGuard uses barrier that's unsuitable for general use case, and results in multiple threads entering critical sections due to premature barrier optimizations. This fixes it.
• Additionally, the 1ms thread sleep isn't helpful, which on platforms like Linux which default to higher resolution thread quantums might be palatable, but will have heavy regressive behaviour on platforms like Windows which default to 15ms quantum slices.
• Introduce AtomicMutex that progressively spins, yields and sleeps to remove this, while also working with regular unique_lock patterns.
• The goal is to remove CLockFreeGuard entirely in subsequent PRs and replace it with standard patterns of std::unique_lock coupled with AtomicMutex

[Jouzo]

Suggested change

	// We could theoritically use CAS here to prevent the additional write, but
	// but requires loop on weak, or using strong. Simpler to just use an exchange for
	// for now, since all ops are seq_cst anyway.
	// We could theoritically use CAS here to prevent the additional write
	// but requires loop on weak, or using strong. Simpler to just use an exchange
	// for now, since all ops are seq_cst anyway.