perf: use torch.embedding for type embedding (deepmodeling#4747)

The backward step of indexing operation is costly on GPU. Using
dedicated `torch.embedding` mitigates this problem.

<details><summary>Profiling results</summary>
<p>
Before: 32ms

<img width="612" alt="image"
src="https://github.com/user-attachments/assets/6e2a4de1-433a-4b6a-8b59-a8458d66897c"
/>
---

After: 0.5ms

<img width="334" alt="image"
src="https://github.com/user-attachments/assets/199ac925-8382-4a43-a6bb-584bb60159b2"
/>

</p>
</details> 



<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

- **Refactor**
- Updated the embedding lookup mechanism in the model, potentially
improving how embeddings are retrieved internally. No changes to the
user interface or method signatures.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Author: caic99

deepmd/pt/model/network/network.py

@@ -286,7 +286,7 @@ def forward(self, atype):
         type_embedding:
 
         """
-        return self.embedding(atype.device)[atype]
+        return torch.embedding(self.embedding(atype.device), atype)
 
     def get_full_embedding(self, device: torch.device):
         """