fix huggingface models input issue in torchbench (#126579)

Summary:
Fixes #2263.

According to https://github.com/pytorch/pytorch/blob/main/benchmarks/dynamo/common.py#L509, example_inputs are formatted as dictionaries for HuggingFace models. However, this forward_pass function passes all inputs to mod with *, which may only pass the input_ids key in HuggingFace model's example inputs.

To reproduce, run the following command.
```bash
python pytorch/benchmarks/dynamo/torchbench.py --performance --inference -dcuda --only=hf_Bert --output=torchbench_inference.csv
```

X-link: pytorch/pytorch#126579
Approved by: https://github.com/xuzhao9

Reviewed By: DanilBaibak

Differential Revision: D57619425

Pulled By: FindHao

fbshipit-source-id: 9074e227bab949c83768092cf019dfd590fe947b

Author: FindHao

userbenchmark/dynamo/dynamobench/torchbench.py

@@ -423,13 +423,19 @@ def compute_loss(self, pred):
 
     def forward_pass(self, mod, inputs, collect_outputs=True):
         with self.autocast(**self.autocast_arg):
-            return mod(*inputs)
+            if isinstance(inputs, dict):
+                return mod(**inputs)
+            else:
+                return mod(*inputs)
 
     def forward_and_backward_pass(self, mod, inputs, collect_outputs=True):
         cloned_inputs = clone_inputs(inputs)
         self.optimizer_zero_grad(mod)
         with self.autocast(**self.autocast_arg):
-            pred = mod(*cloned_inputs)
+            if isinstance(clone_inputs, dict):
+                pred = mod(**cloned_inputs)
+            else:
+                pred = mod(*cloned_inputs)
             loss = self.compute_loss(pred)
         self.grad_scaler.scale(loss).backward()
         self.optimizer_step()