Declare free-threaded safety for _scalene_unwind module (#1065)

The _scalene_unwind native extension (native_unwind.cpp) was the only one
of Scalene's three C extension modules that did not call
PyUnstable_Module_SetGIL(m, Py_MOD_GIL_NOT_USED) in its init. On
free-threaded builds (3.13t/3.14t) importing a module that has not declared
free-threaded safety makes CPython silently re-enable the GIL and emit a
RuntimeWarning. Because Scalene always imports _scalene_unwind, the GIL was
restored on every free-threaded run -- so Scalene never actually ran
free-threaded, adding startup overhead and serializing threads.

pywhere.cpp and get_line_atomic.cpp already make this call; this brings
_scalene_unwind in line with them.

Verified on Linux (3.13t and 3.14t): after the fix sys._is_gil_enabled()
stays False after importing scalene, the RuntimeWarning is gone, the memory
attribution suite passes, and the free-threaded parity test shows the
contention ratio drop below 1.0 (real parallel scaling) instead of being
masked by a re-enabled GIL.

Author: emeryberger

src/source/native_unwind.cpp

@@ -806,6 +806,13 @@ PyModuleDef moduledef = {
 extern "C" PyObject* PyInit__scalene_unwind(void) {
   PyObject* m = PyModule_Create(&moduledef);
   if (!m) return nullptr;
+#ifdef Py_GIL_DISABLED
+  // Declare free-threaded safety so importing this module does not silently
+  // re-enable the GIL on free-threaded (3.13t/3.14t) builds. Without this,
+  // CPython restores the GIL with a RuntimeWarning, so Scalene never actually
+  // runs free-threaded -- matches pywhere.cpp and get_line_atomic.cpp.
+  PyUnstable_Module_SetGIL(m, Py_MOD_GIL_NOT_USED);
+#endif
   PyModule_AddIntConstant(m, "available", SCALENE_UNWIND_AVAILABLE);
   // Expose the per-call cap so callers (and tests) read the single
   // source of truth instead of hard-coding the value.