cuda.core: harden graph user-object destructor during Python shutdown

cuda_core/cuda/core/graph/_utils.pxd

@@ -7,7 +7,7 @@ from cuda.bindings cimport cydriver
 
 cdef bint _is_py_host_trampoline(cydriver.CUhostFn fn) noexcept nogil
 
-cdef void _py_host_destructor(void* data) noexcept with gil
+cdef void _py_host_destructor(void* data) noexcept nogil
 
 cdef void _attach_user_object(
     cydriver.CUgraph graph, void* ptr,

cuda_core/cuda/core/graph/_utils.pyx

@@ -14,15 +14,42 @@ from cuda.core._utils.cuda_utils cimport HANDLE_RETURN
 
 
 cdef extern from "Python.h":
+    int Py_IsInitialized() nogil
     void _py_decref "Py_DECREF" (void*)
 
+cdef extern from *:
+    """
+    #include <Python.h>
+
+    #if PY_VERSION_HEX < 0x030D0000
+    extern int _Py_IsFinalizing(void);
+    #endif
+
+    static inline int _py_runtime_is_finalizing(void) {
+    #if PY_VERSION_HEX >= 0x030D0000
+        return Py_IsFinalizing();
+    #else
+        return _Py_IsFinalizing();
+    #endif
+    }
+    """
+    int _py_runtime_is_finalizing() nogil
+
 
 cdef void _py_host_trampoline(void* data) noexcept with gil:
     (<object>data)()
 
 
-cdef void _py_host_destructor(void* data) noexcept with gil:
-    _py_decref(data)
+cdef void _py_host_destructor(void* data) noexcept nogil:
+    if data == NULL:
+        return
+
+    # Leak once shutdown has started or completed.
+    if not Py_IsInitialized() or _py_runtime_is_finalizing():
+        return
+
+    with gil:
+        _py_decref(data)
 
 
 cdef bint _is_py_host_trampoline(cydriver.CUhostFn fn) noexcept nogil: