[inductor][cpu]disable pointwise_cat on CPU (pytorch#116313)

We observed negative performance impact of pointwise_cat optimization on CPU so disabled it. We will revisit this later after enabling vectorization on index_expr.

This PR fix the following three regression issues:
pytorch#115827
pytorch#112139
pytorch#114495

and cause performance regression of pytorch_unet again. Related issue: pytorch#115343

Pull Request resolved: pytorch#116313
Approved by: https://github.com/jgong5, https://github.com/leslie-fang-intel, https://github.com/eellison

test/inductor/test_cpu_repro.py

@@ -2708,7 +2708,7 @@ def fn(x, y):
         y = torch.randn(32, 120)
         metrics.reset()
         self.common(fn, (x, y))
-        assert metrics.generated_cpp_vec_kernel_count == 1
+        assert metrics.generated_cpp_vec_kernel_count == 3
 
     def test_expr_vec_non_contiguous(self):
         def fn(x):

torch/_inductor/lowering.py

@@ -1092,7 +1092,12 @@ def should_lower_cat_input(x) -> bool:
 
         return False
 
-    if len(inputs) <= config.max_pointwise_cat_inputs:
+    # TODO: We observed negative performance impact of pointwise_cat optimization on CPU so disabled it.
+    #             We will revisit this later after enabling vectorization on index_expr.
+    if (
+        len(inputs) <= config.max_pointwise_cat_inputs
+        and inputs[0].get_device().type != "cpu"
+    ):
         pointwise_uses = all(is_pointwise_use(use) for use in V.current_node.users)
         all_pointwise_inputs = all(should_lower_cat_input(inp) for inp in inputs)
         any_pointwise_inputs = any(should_lower_cat_input(inp) for inp in inputs)