[BUILD] Upgrade torch-npu to 2.5.1 (#661)

### What this PR does / why we need it?
The torch-npu 2.5.1 are published:
https://pypi.org/project/torch-npu/2.5.1/
It's time to remove all torch-npu dev version from vllm-ascend code base

### Does this PR introduce _any_ user-facing change?
Yes, using torch-npu 2.5.1

### How was this patch tested?
- [ ] CI passed
- [ ] Manually test
- [ ] Grep all `dev2025`

---------

Signed-off-by: Yikun Jiang <yikunkero@gmail.com>

vllm_ascend/models/qwen2_vl.py

@@ -86,7 +86,7 @@ class CustomQwen2VisionAttention(Qwen2VisionAttention):
 
         context_layer = torch.torch.empty_like(q)
 
-        # operator requires pta version >= 2.5.1.dev20250226
+        # operator requires pta version >= 2.5.1
         torch_npu._npu_flash_attention_unpad(
             query=q,
             key=k,