Don't Decompose Hardswish (pytorch#12360)

Summary:

Investigating MV3, I noticed that hardswish was getting decomposed into many little ops. This become annoying because it injected unnecessary transposes, and also wasn't being quantized. I didn't realize that this was being decomposed. After doing some investigating, it looks like this can greatly improve our MV3 Performance for both Quantized and FP32 models. Some benchmarks here.

|                | Before Hardwish Decomp | After Hardswish Decomp  | Latency Reduction |
|----------------|------------------------|------------------------|-------------------|
| Macbook (FP32) | [13.3685]((https://www.internalfb.com/phabricator/paste/view/P1859573931))                | [8.451](https://www.internalfb.com/phabricator/paste/view/P1859573328)                  |36%               |
| Macbook (QS8)  | [16.0361](https://www.internalfb.com/phabricator/paste/view/P1859609658)                | [4.914](https://www.internalfb.com/phabricator/paste/view/P1859610252)                  |69%               |
| S24 (FP32)     | [56.885](https://www.internalfb.com/intern/paste/P1859603500)                 | [41.9638](https://www.internalfb.com/intern/paste/P1859603738)                |26%               |
| S24 (QS8)      | [56.1718](https://www.internalfb.com/intern/paste/P1859615896)                | [40.2096](https://www.internalfb.com/intern/paste/P1859615683/)                |40%               |

Reviewed By: cccclai

Differential Revision: D77765129

Author: mcr229

backends/xnnpack/partition/config/generic_node_configs.py

@@ -399,6 +399,9 @@ class HardswishConfig(GenericNodePartitionerConfig):
 
     def supported_precision_types(self) -> List[ConfigPrecisionType]:
         return [ConfigPrecisionType.FP32]
+    
+    def get_original_aten(self) -> Optional[torch._ops.OpOverload]:
+        return torch.ops.aten.hardswish.default
 
 
 class LeakyReLUConfig(GenericNodePartitionerConfig):