Refactor generalized eigensolvers to avoid explicit Cholesky inverse

## Background

Several code paths solve generalized Hermitian eigenproblems with an overlap matrix using Cholesky orthogonalization. For example, `dptb/nn/dftb/dftb_scc.py::SKSCC.eigh_solver` currently computes:

```python
chklowt = torch.linalg.cholesky(overlap_mat)
chklowtinv = torch.linalg.inv(chklowt)
H_mat = chklowtinv @ H_mat @ chklowtinv.transpose(1, 2).conj()
eigval, eigvec = torch.linalg.eigh(H_mat)
eigvec = chklowtinv.transpose(1, 2).conj() @ eigvec
```

This is mathematically valid, but explicitly inverting the triangular Cholesky factor is less efficient and less numerically stable than using `torch.linalg.solve_triangular`.

## Proposed direction

Introduce a shared helper for generalized Hermitian eigenproblems, e.g.

```python
solve_generalized_eigh(H, S=None) -> eigval, eigvec
```

with a single documented convention for eigenvector shape, likely `(nk, norb, nstate)`.

The helper should replace duplicated logic in:

- `dptb/nn/dftb/dftb_scc.py::SKSCC.eigh_solver`
- `dptb/nn/energy.py::Eigh`
- potentially `dptb/nn/energy.py::Eigenvalues`

## Important details

For `S = L L†`, the standard transformation is:

```text
H' = L⁻¹ H L⁻†
H' c = E c
psi = L⁻† c
```

When replacing explicit inverse with triangular solves, the right-side transform and eigenvector back-transform need careful treatment. A naïve patch may preserve eigenvalues while returning incorrectly transformed eigenvectors, which can silently affect Mulliken charges and SCC workflows.

## Suggested tests

Add tests that compare the new helper against the current inverse-based implementation for:

- batched real Hermitian matrices
- batched complex Hermitian matrices
- eigenvalue agreement
- generalized eigenvector orthonormality: `eigvec.conj().T @ S @ eigvec == I`
- consistency with existing SCC / `Eigh` eigenvector shape conventions

## Notes

This is not currently blocking PR #330. It is a valid numerical/performance improvement, but should be handled as a focused follow-up with tests rather than as a quick inline replacement.