Update on "Better handing of Autograd+Fork errors."

Fixes: #32835
Fixes: #5834

Can not combine with CUDA's implementation as each of them requires individual `std::once_flag` as well as different `forked_autograd_child` functions. CUDA version relays to python module, autograd uses TORCH_CHECK to report error to python and cpp.

Differential Revision: [D20144024](https://our.internmc.facebook.com/intern/diff/D20144024)

[ghstack-poisoned]

test/test_multiprocessing.py

@@ -356,15 +356,15 @@ def test_inherit_tensor(self):
         p.join(1)
         self.assertEqual(t, torch.ones(5, 5) * 3, 0)
 
-    @unittest.skipIf(IS_WINDOWS, "Test need to use fork multiprocessing")
+    @unittest.skipIf(IS_WINDOWS, "Test needs to use fork multiprocessing")
     def test_autograd_errors(self):
         ctx = mp.get_context('fork')
         simple_autograd_function()
         with self.assertRaisesRegex(RuntimeError, r'Unable to handle autograd'):
             with ctx.Pool(3) as pool:
                 pool.map(simple_autograd_function, [1, 2, 3])
 
-    @unittest.skipIf(NO_MULTIPROCESSING_SPAWN, "Test need to use spawn multiprocessing")
+    @unittest.skipIf(NO_MULTIPROCESSING_SPAWN, "Test needs to use spawn multiprocessing")
     def test_autograd_fine_with_spawn(self):
         ctx = mp.get_context('spawn')
         simple_autograd_function()