Fix autograd thread crash with python-3.9 (#50998)

Summary:
Update pybind repo to include `gil_scoped_acquire::disarm()` methods
In python_engine allocate scoped_acquire as unique_ptr and leak it if engine is finalizing for Python-3.9+

Fixes https://github.com/pytorch/pytorch/issues/50014 and https://github.com/pytorch/pytorch/issues/50893

Pull Request resolved: https://github.com/pytorch/pytorch/pull/50998

Reviewed By: ezyang

Differential Revision: D26038314

Pulled By: malfet

fbshipit-source-id: 035411e22825e8fdcf1348fed36da0bc33e16f60

torch/csrc/autograd/python_engine.cpp

@@ -48,6 +48,10 @@ Engine& PythonEngine::get_python_engine() {
   return engine;
 }
 
+#if PY_MAJOR_VERSION == 3 && PY_MINOR_VERSION >= 9
+#define IS_PYTHON_3_9_PLUS
+#endif
+
 void PythonEngine::thread_init(int device, const std::shared_ptr<ReadyQueue>& ready_queue, bool should_increment) {
   // Increment thread usage count before acquiring the GIL
   if (should_increment) {
@@ -56,7 +60,11 @@ void PythonEngine::thread_init(int device, const std::shared_ptr<ReadyQueue>& re
   // Create a PyThreadState, but release the GIL. This lets pybind11::gil_scoped_acquire calls
   // inside thread_main acquire the GIL without having to create a new
   // PyThreadState each time.
+#ifdef IS_PYTHON_3_9_PLUS
+  auto gil = std::make_unique<pybind11::gil_scoped_acquire>();
+#else
   pybind11::gil_scoped_acquire gil;
+#endif
   pybind11::gil_scoped_release no_gil;
   Engine::thread_init(device, ready_queue, false);
 
@@ -64,6 +72,15 @@ void PythonEngine::thread_init(int device, const std::shared_ptr<ReadyQueue>& re
     // Decrement the count during shutdown if we incremented earlier.
     decrement_non_reentrant_thread_count();
   }
+
+#ifdef IS_PYTHON_3_9_PLUS
+  // Do not call PyEval_RestoreThread, PyThreadState_[Clear|DeleteCurrent] if runtime is finalizing
+  if (_Py_IsFinalizing()) {
+    no_gil.disarm();
+    // TODO: call disarm rather than leak gil_scoped_acquired once PyThreadState_Clear can safely be called from finalize
+    gil.release();
+  }
+#endif
 }
 
 void PythonEngine::thread_on_exception(