Future atomic operation cleanup

[Roland McGrath]

I don't see a problem with conditional code based on available compiler
version--that is, conditional code within the atomic.h implementations
themselves.  What I take to be the main thrust of Torvald's suggestion is
that we get ASAP to a situation where all the code outside of actual
atomic.h implementations themselves is using a new atomic.h API that has
semantics and signatures identical to, and names similar to, the
C11-supporting builtins.  That enables us to clean up, fix, and optimize
all the actual uses of atomics throughout the codebase, without waiting
until we can rely on requiring newer compiler versions throughout.

I tend to think it is still too early to mandate 4.7 or later.  I think we
can in the 2.21 cycle move up to 4.6 as the minimum (and 4.8 or
4.9 as the recommended version).  But we should discuss that further.