Match Torch speed for sum reduction on M1 (#1187)

* Add additional kernel when reducing multiple dimensions at once.

* Faster for smaller inputs

* Whitespace and naming

* Cleaner, guard for Metal only, and max 1 split rather than N

* Draft of different approach

* One additional kernel call for this test (as expected)

test/external/external_test_opt.py

@@ -63,7 +63,7 @@ class TestInferenceMinKernels(unittest.TestCase):
     for p in get_parameters(model): p.assign(np.zeros(p.shape, dtype=p.dtype.np))
     img = Tensor.randn(1, 3, 224, 224)
     # TODO: this seems very high
-    with CLCache(115):
+    with CLCache(116):
       model.forward(img).realize()
 
   def test_resnet(self):