pd: revert einsum to matmul for paddle backend (#4768)

revert einsum to matmul for paddle backend as einsum may not trigger
infermeta error.

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

## Summary by CodeRabbit

- **Refactor**
- Improved the internal computation method for certain tensor operations
to enhance performance and maintainability. No changes to user-facing
features or outputs.

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

Signed-off-by: HydrogenSulfate <490868991@qq.com>

Author: HydrogenSulfate

deepmd/pd/model/model/transform_output.py

@@ -82,9 +82,7 @@ def task_deriv_one(
     assert extended_force is not None
     extended_force = -extended_force
     if do_virial:
-        extended_virial = paddle.einsum(
-            "...ik,...ij->...ikj", extended_force, extended_coord
-        )
+        extended_virial = extended_force.unsqueeze(-1) @ extended_coord.unsqueeze(-2)
         # the correction sums to zero, which does not contribute to global virial
         if do_atomic_virial:
             extended_virial_corr = atomic_virial_corr(extended_coord, atom_energy)