Register and unload in a python atexit handler. (#53)

This works around some nasty thread exit behavior from python, described in more detail at python/cpython#87135 .
This behavior was the source of moolib crashes and deadlocks on exit.
Luckily, we can essentially unload moolib in an atexit handler - this waits for any callbacks etc to finish, and prevents any more callbacks from being issued. Any python Rpc objects will still be valid, but closed (any operations on them will silently fail).
CPython will also then not have an opportunity to exit our async threads, which would result in deadlocks.

src/moolib.cc

@@ -1503,13 +1503,20 @@ void set_log_level(py::object level) {
   }
 }
 
+static bool moduleUnloaded = false;
+
 PYBIND11_MODULE(_C, m) {
   moduleIncRef();
-  rpc::moolibModule = m.ptr();
   m.add_object("_cleanup", py::capsule([] {
-                 moduleUnload();
-                 rpc::moolibModule = nullptr;
+                 if (!moduleUnloaded) {
+                   moduleUnload();
+                 }
                }));
+  py::module::import("atexit").attr("register")(py::cpp_function([] {
+    if (!moduleUnloaded) {
+      moduleUnload();
+    }
+  }));
 
   // numpy forks during import, and pybind can trigger a numpy import
   // during serialization (just for checking if the input is a numpy array).