CPython str.find absent-needle slowdown traced to compiled skip step

Absent-needle str.find can be about 3-5x slower on some CPython builds because the skip step in Objects/stringlib/fastsearch.h compiles to a conditional move instead of a branch. On 3.13.5, adding an empty __asm__ volatile("") in that step took a Clang build's find("zq") from about 97us to 28us on arm64 and a PGO+LTO build's 2-char absent search from 71us to 17us on x86, with JSON roughly unchanged. Without PGO, GCC is slow too; GCC with PGO is fast and Clang stayed slow in every setup tried, so this is not simply a compiler-brand issue. On x86 the barrier did not help a 5-char absent search and made rfind slower, so treat it as a diagnostic rather than a fix. Separately, emoji-JSON penalties scale with PGO training: adding non-ASCII text to training cut Clang's ratio from 2.29x to 1.54x on x86.