MY_RELAX_CPU(): On GCC and compatible compilers (including clang and
its derivatives), let us use a null inline assembler block as the
fallback. This should benefit s390x and LoongArch, for example.
Also, let us remove the generic fallback block that does exactly the
opposite of what this function aims to achieve: avoid hogging the
memory bus so that other threads will have a chance to let our spin
loop to proceed.
On RISC-V, we will use __builtin_riscv_pause() which is a valid
instruction encoding in all ISA versions according to
https://gcc.gnu.org/pipermail/gcc-patches/2021-January/562936.html
