Consider using __atomic* builtins instead of __sync* ones: the first support 64-bit atomics on all platforms

[barracuda156]

src/tcpkali_atomic.h uses old __sync* built-ins, which only support 32-bit atomics on 32-bit platforms, and then we get:

#warning \
    "This compiler does not have 64-bit compare_and_swap, results might be broken"

Instead, if the code will be changed to use __atomic* built-ins (which is what gcc upstream recommends), 64-bit atomics will be available on all platforms (32-bit ones may need explicit linking to libatomic).