feat: add CUDA-to-NPU monkey patches for GPUModelRunner compatibility

vllm_npu/__init__.py

@@ -9,4 +9,10 @@ class name of the platform implementation.
 
 def register():
     """Return the fully-qualified name of the NPU platform class."""
+    # Apply CUDA→NPU compatibility patches early so that any code
+    # referencing torch.cuda.Stream / Event / etc. will transparently
+    # be redirected to the torch.npu equivalents.
+    from vllm_npu.cuda_compat import _patch_cuda_to_npu
+    _patch_cuda_to_npu()
+
     return "vllm_npu.platform.NPUPlatform"