[MISC] Clean up torch_npu (#688)

torch_npu 2.5.1 support autoload now. This patch does:
1. remove useless torch_npu import
2. replace `torch_npu.npu` to `torch.npu`.

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm_ascend/patch/worker/patch_0_8_4/patch_metrics.py

@@ -18,7 +18,6 @@
 from typing import Callable, Optional, Union
 
 import torch
-import torch_npu
 from vllm.spec_decode.metrics import (AsyncMetricsCollector,
                                       SpecDecodeWorkerMetrics)
 
@@ -36,7 +35,7 @@ def init_tensors(self,
     if isinstance(device_type, torch.device):
         device_type = device_type.type
     if device_type == 'npu':
-        self._copy_stream = torch_npu.npu.Stream()
+        self._copy_stream = torch.npu.Stream()
 
 
 def maybe_collect_rejsample_metrics(