enable fx graph cache on torchbench (#128239)

Summary:
We've already enabled for timm and huggingface, but we had failures saving cache entries for moco. It looks like pytorch/pytorch#128052 has fixed that issue, so we can enable for torchbench.

X-link: pytorch/pytorch#128239
Approved by: https://github.com/oulgen

Reviewed By: clee2000

Differential Revision: D58394057

Pulled By: masnesral

fbshipit-source-id: 58a226babdcd04a3bb14118e7a555558c56bc9ef

userbenchmark/dynamo/dynamobench/torchbench.py

@@ -25,6 +25,10 @@
 # We are primarily interested in tf32 datatype
 torch.backends.cuda.matmul.allow_tf32 = True
 
+# Enable FX graph caching
+if "TORCHINDUCTOR_FX_GRAPH_CACHE" not in os.environ:
+    torch._inductor.config.fx_graph_cache = True
+
 
 def _reassign_parameters(model):
     # torch_geometric models register parameter as tensors due to