fix(pt): detach computed descriptor tensor to prevent OOM (deepmodeli…

…ng#4547)

Fix deepmodeling#4544.

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

## Summary by CodeRabbit

- **Optimization**
- Improved descriptor handling in the atomic model to enhance
computational efficiency and memory management during backpropagation.

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

Signed-off-by: Jinzhe Zeng <[email protected]>

deepmd/pt/model/atomic_model/dp_atomic_model.py

@@ -244,7 +244,7 @@ def forward_atomic(
         )
         assert descriptor is not None
         if self.enable_eval_descriptor_hook:
-            self.eval_descriptor_list.append(descriptor)
+            self.eval_descriptor_list.append(descriptor.detach())
         # energy, force
         fit_ret = self.fitting_net(
             descriptor,