Release pool unsoundness

[Alexander-N]

Follow up for #720 where NameErrors and SIGSEGVs were found. The error can be observed with two tests running in parallel (might need a few tries):

#[test]
fn test_release_race1() {
    let gil = Python::acquire_gil();
    let py = gil.python();
    println!("make ob1");
    let ob1 = py.eval("object()", None, None).unwrap();
    py.import("datetime").unwrap();
    py_run!(py, ob1, "print('ob1' + str(ob1))");
}

#[test]
fn test_release_race2() {
    let gil = Python::acquire_gil();
    let py = gil.python();
    println!("make ob2");
    let ob2 = py.eval("object()", None, None).unwrap();
    py.import("datetime").unwrap();
    py_run!(py, ob2, "print('ob2' + str(ob2))");
}

What might be happening is:

• ob1 gets put into the release pool
• the import releases the GIL temporarily
• test_release_race2 makes progress and ob2 gets put into the release pool
• test_release_race1 finishes and with dropping of it's GILGuard ob2 instead of ob1 is popped from the relase pool
• test_release_race2 fails because ob2 has been freed

[davidhewitt]

Now that #887 is merged we should have resolved the thread safety issues here 🎉