Add lru_cache for node comparison (#148)

Summary:
Pull Request resolved: #148

cache the args to optimize the number of calls. For `is_same_node` function, we're expecting the same results each time it's called with the same inputs.

For mv2,
```
buck run mode/dev-nosan  //executorch/examples/backend:xnnpack_examples -- -m mv2 -q -d
```
The total number of nodes in mv2 graph after quantize is 520.

Before the change:
mv2: The number of call to `is_same_node` is 6388036 🙈
mv2: to_backend call time: 30.79 second (including dumping the logs)
mv3: takes forever

After the change:
mv2: The number of call to `is_same_node` is 520. It's more reasonable
mv2: to_backend call time: 10.15 second (including dumping the logs)
mv3: `to_backend` call takes 16 second.

Reviewed By: digantdesai, mcr229

Differential Revision: D48708658

fbshipit-source-id: 8585c417d6d7c27bbacda563c0a7a7a764f0674e

Author: cccclai

exir/backend/utils.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+from functools import lru_cache
 from typing import Iterable, List, Tuple
 
 import torch
@@ -16,6 +17,7 @@
 T_DQuantPerTensor = exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default
 
 
+@lru_cache(maxsize=128)
 def is_same_node(
     node_left: Iterable[torch.fx.Node],
     node_right: Iterable[torch.fx.Node],
@@ -39,8 +41,6 @@ def is_same_node(
         if len(list(node_left)) != len(list(node_right)):
             return False
         for n_left, n_right in zip(node_left, node_right):
-            # pyre-fixme[6]: For 1st argument expected `Iterable[Node]` but got `Node`.
-            # pyre-fixme[6]: For 2nd argument expected `Iterable[Node]` but got `Node`.
             if not is_same_node(n_left, n_right):
                 return False
     return True