Property prediction model training fails when training data is much similar

### Summary

A researcher trains a d33(piezoelectric coefficient) prediction model with 2 training data using property prediction head. The two data is perovskite-type(ABO3). The composition(Pb1000Ti340Nb400Mg140In120O3000) and the spatial arrangement of atoms of the two data are completely the same. The only difference between the two data is that Ti,Nb,Mg,In randomly occupy the B sites in 10✖️10✖️10 perovskite supercells. The two pictures shows the x-direction projection of B sites in the two data. I use `StructureMatcher` of `pymatgen` to confirm that the two structure can not match. The label of the two data is 411 and 587, respectively.

I use the settings of DPA-3.1-3M and train 1000 steps(500 epoch), but find there is no precision, the `dp test` result is
```
# d33_database/1 - 0: data_property pred_property
4.111499938964843750e+02 4.762894698179545685e+02
# d33_database/2 - 0: data_property pred_property
5.865300292968750000e+02 4.762894713410120744e+02
```
It seems that dp model can not distinguish the two similar structure.

The reproduce `input.json` and data file are:
[input.json](https://github.com/user-attachments/files/23020817/input.json)
[d33_database.tgz](https://github.com/user-attachments/files/23020837/d33_database.tgz)

### DeePMD-kit Version

3.1.0

### Backend and its version

pytorch

### Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

Python version: 3.12.11
CUDA Version: 12.4

### Details

See above

Start by running the linked input.json with d33_database.tgz under DeePMD-kit 3.1.0 and the PyTorch backend, then compare the reported dp test predictions for both structures. Read the property-prediction training entry points and check how the DPA-3.1-3M settings handle these two samples; done means identifying and documenting the cause of the indistinguishable predictions or confirming the reproducible defect.

Written by the indexing model from the issue text.