Stabilise vibrational eigenvalue filtering

[harryswift01]

Summary

This PR improves the numerical stability of vibrational entropy calculations by using the symmetric covariance eigensolver and applying scale-aware eigenvalue filtering.

The change targets intermittent regression failures caused by near-zero covariance eigenmodes being handled inconsistently across platforms.

Changes

Use symmetric covariance eigensolver:

• Symmetrise covariance matrices before eigendecomposition.
• Replace the general eigenvalue solver with the symmetric/Hermitian eigensolver.
• Validate that vibrational entropy inputs are square matrices.

Improve eigenvalue filtering:

• Replace complex-value filtering with real finite-value filtering.
• Exclude non-positive and near-zero eigenvalues using a scale-aware threshold.
• Log how many eigenvalues are excluded and the tolerance used.

Make flexible-dihedral handling deterministic:

• Convert eigenvalues to a real NumPy array before adjustment.
• Select the largest eigenvalues using deterministic NumPy sorting.
• Guard against flexible counts larger than the available eigenvalue count.

Impact

• Reduces platform-sensitive vibrational entropy differences in CI.
• Handles rigid-body/null covariance modes more robustly.
• Improves numerical correctness for symmetric covariance matrices.
• May require regression baselines to be regenerated because this changes eigenvalue filtering behaviour.