Stabilise vibrational regression calculations

[harryswift01]

Summary

This PR improves numerical stability and determinism in the vibrational entropy regression path.

The changes focus on using the appropriate symmetric eigensolver for covariance matrices and ensuring group traversal/indexing is deterministic. This targets intermittent CI-only regression failures where vibrational entropy components differed between otherwise equivalent runs.

Changes

Stabilise covariance eigendecomposition

• Replace the general eigenvalue solver with the symmetric/Hermitian eigensolver for covariance matrices.
• Explicitly symmetrise covariance matrices before eigenvalue calculation.
• Avoid unnecessary complex eigenvalue paths for real symmetric covariance data.

Enforce deterministic group handling

• Build group-to-index mappings using sorted group IDs.
• Iterate over water groups in sorted order.
• Iterate over neighbour/symmetry groups in sorted order.

Impact

• Improves reproducibility of vibrational entropy calculations across CI and local environments.
• Reduces sensitivity to platform-specific LAPACK/BLAS behaviour.
• Prevents group-index mappings from depending on dictionary insertion order.
• Does not change the public API or regression workflow structure.