Fix static data initialization deadlock on GIL (#34505)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34505
A thread could hold GIL when calling PythonRpcHandler::getInstance(),
meantime another thread could have been doing static data
initialization by calling `new PythonRpcHandler()`, inside of which GIL is
also required. Static data initialization is thread-safe, so the thread
holding the GIL will wait for the other thread to finish static data
initializating before going forward. Because the initialization can't
proceed without GIL, there is a deadlock. We ask the calling thread to
release GIL to avoid this situation.
ghstack-source-id: 99893858
Test Plan:
```
buck test mode/dev-nosan //caffe2/test/distributed/rpc:dist_autograd_spawn -- 'test_backward_simple_script_call \(test_dist_autograd_spawn\.DistAutogradTestWithSpawn\)' --stress-runs 100
```
Differential Revision: D7490489
fbshipit-source-id: 76f63cc7bedf088d3dbff288f53aa0bd33749255
