gh-157194: Read the opcode atomically in the specializer

[iamsharduld]

set_opcode() writes instr->op.code with _Py_atomic_compare_exchange_uint8() and unspecialize() reads it with FT_ATOMIC_LOAD_UINT8_RELAXED(), the convention introduced with the (un)specialization helpers in gh-115999 (9ce4fa0). Two later comparisons read the same byte directly:

• _Py_Specialize_Resume(), added by 3d0824a (Trace starting from RESUME in the JIT #127958)
• _Py_Specialize_BinaryOp(), added by 3893a92 (Specialize long tail of binary operations using a table. #100239)

In a free-threaded build each thread normally gets its own copy of the bytecode, so specialization never races. With -X tlbc=0 the copies are disabled and the bytecode is shared, and those plain reads race with the atomic write. ThreadSanitizer reported exactly that on the CI job for test_thread_local_bytecode (report quoted in the issue): a read at specialize.c:2957 against a "previous atomic write" from set_opcode() via unspecialize().

The change makes both reads use the same relaxed atomic load. The logic is unchanged, and on the default build FT_ATOMIC_LOAD_UINT8_RELAXED expands to the plain access, so only the free-threaded build is affected.

Verification

On a --disable-gil --with-thread-sanitizer --with-pydebug build (clang 20, Linux aarch64), with a script that has several threads and the main thread call the same freshly created function repeatedly under -X tlbc=0:

	data races reported
before	3, every one of them _Py_Specialize_Resume Python/specialize.c:2957:43 — the same file, line and column as the CI report
after	0, in three consecutive runs

test_thread_local_bytecode passes 10 out of 10 runs on that build, and test_opcache, test_capi, test_code, test_monitoring, test_sys_settrace and test_generators pass under TSan (2244 tests). On macOS, test_thread_local_bytecode, test_opcache, test_monitoring and test_capi pass on a free-threaded debug build, and test_opcache, test_dis, test_monitoring, test_generators, test_sys_settrace and test_capi pass on the default build.

• Issue: Data race on instr->op.code in _Py_Specialize_Resume() with tlbc disabled (TSan CI failure in test_thread_local_bytecode) #157194