Comparing one function across two builds: mask RIP-relative operands

To check whether two builds (e.g. independently trained PGO binaries) generated the same code for a function, raw bytes are useless: rel32 call/jump targets and RIP-relative data displacements change whenever the function or its callees move. On two gcc 13.3 PGO+LTO CPython 3.14.0 builds, only 743 of 10086 same-size text symbols were raw-identical. Normalize instead: get start/size from nm -S, objdump -d --no-show-raw-insn --start-address --stop-address, rewrite <hex> <sym> targets as function-relative (inside) or EXT (outside), strip # comments, and replace 0x..(%rip) with a placeholder. Then the hot search kernels compared identical and only their start address mod 64 differed. Masking targets alone wasn't enough: _Py_bytes_count still differed until RIP displacements were masked.