skip special tokens (#715)

it should skip special tokens here. just like trl do
https://github.com/huggingface/trl/blob/fc2b041b58f6fbe766dceaec819bc5a8f9d209da/trl/trainer/grpo_trainer.py#L597


if `skip_special_tokens=False`,  completion 

```
<think>...</think><answer>....</answer>
```

will be decoded as things such as
```
<think>...</think><answer>....</answer><|im_end|><|endoftext|>
```

which will render typical `format_reward_func` mismatch

```python
r"^<think>.*?</think>\s*<answer>.*?</answer>$"
```

verl/workers/reward_manager/naive.py

@@ -43,8 +43,8 @@ class NaiveRewardManager:
             valid_response_ids = response_ids[:valid_response_length]
 
             # decode
-            prompt_str = self.tokenizer.decode(valid_prompt_ids)
-            response_str = self.tokenizer.decode(valid_response_ids)
+            prompt_str = self.tokenizer.decode(valid_prompt_ids, skip_special_tokens=True)
+            response_str = self.tokenizer.decode(valid_response_ids, skip_special_tokens=True)
 
             ground_truth = data_item.non_tensor_batch['reward_model']['ground_truth']
 
@@ -88,8 +88,8 @@ class NaiveRewardManager:
             valid_response_ids = response_ids[:valid_response_length]
 
             # decode
-            prompt_str = self.tokenizer.decode(valid_prompt_ids)
-            response_str = self.tokenizer.decode(valid_response_ids)
+            prompt_str = self.tokenizer.decode(valid_prompt_ids, skip_special_tokens=True)
+            response_str = self.tokenizer.decode(valid_response_ids, skip_special_tokens=True)
 
             ground_truth = data_item.non_tensor_batch['reward_model']['ground_truth']