Add noexcept(false) to destructors for gil_scoped_release

When a thread attempts to acquire the GIL but the Python interpreter has already destructed, Python will attempt to terminate the thread using pthread_exit. In many implementations of pthread_exit, this will trigger a stack unwinding, which will immediately call std::terminate if you are inside a destructor with noexcept(true). Which is the case for the destructor of gil_scoped_release, which will attempt to acquire the GIL on reentry. The net effect of this any code which uses gil_scoped_release and is called from a daemon thread from Python is likely to cause your process to unceremoniously exit.

The fix seems to be quite simple:

diff --git a/include/pybind11/pybind11.h b/include/pybind11/pybind11.h
index 7fa0f0e..eb501a7 100644
--- a/include/pybind11/pybind11.h
+++ b/include/pybind11/pybind11.h
@@ -1954,7 +1954,7 @@ public:
             PYBIND11_TLS_DELETE_VALUE(key);
         }
     }
-    ~gil_scoped_release() {
+    ~gil_scoped_release() noexcept(false) {
         if (!tstate)
             return;
         PyEval_RestoreThread(tstate);

Do you agree with this change? If so I can submit a PR for it. More context: https://github.com/pytorch/pytorch/issues/38228

It would be reasonably simple to produce a repro that doesn't involve PyTorch, please let me know if that would be helpful.

Start in include/pybind11/pybind11.h at the gil_scoped_release destructor and review the issue's proposed exception specification. Check the surrounding GIL and thread-shutdown behavior, then validate the change with a focused reproduction involving interpreter destruction and a daemon thread. Done means the failure no longer terminates the process unexpectedly.