[Code scan] Fix diff_on eigenvalue loss indexing for 2D band tensors

This issue is part of a Codex global repository scan.

Problem:
`EigLoss.forward` asserts that `eig_pred_cut` and `eig_label_cut` are 2D tensors, but the `diff_on` branch indexes them as if they were 3D (`[:, k_diff_i, :]`).

Code references:
https://github.com/deepmodeling/deeptb/blob/86c60c73996f0dd961c3138f2e88424382cb734e/dptb/nnops/loss.py#L188-L193
https://github.com/deepmodeling/deeptb/blob/86c60c73996f0dd961c3138f2e88424382cb734e/dptb/nnops/loss.py#L223-L237

Impact:
Any run with `diff_on=True` raises an indexing error instead of computing the differential eigenvalue loss.

Suggested fix:
Index the k-point dimension consistently for 2D tensors, for example `eig_label_cut[k_diff_i, :] - eig_label_cut[k_diff_j, :]`, with the same correction for predictions and masked variants.

Start in dptb/nnops/loss.py at the referenced EigLoss.forward sections around lines 188-193 and 223-237. Trace the diff_on path and its masked variants for the asserted 2D tensors. Done means diff_on=True no longer raises an indexing error and computes the differential eigenvalue loss consistently for predictions and labels.

Written by the indexing model from the issue text.