Bug in the interpolation with LO-TO splitting

[mesonepigreco]

As also noted in https://github.com/orgs/SSCHAcode/discussions/353,
there is a bug in the LO-TO splitting.

THis is a way to reproduce it systematically.

The two following dynamical matrices represent the same structure, however the rhombohedral present a slight rhombohedral distortion due to stochastic noise in a SSCHA relaxation where the cubic is the ground state.
Also they have a rotation between them.

dyn_cubic_with_born.zip
dyn_rhonbohedral_with_born.zip

If we interpolate using InterpolateMesh, also, effective charges and dielectric tensor have been computed on the similar structures and are almost identical.

The frequencies of the dynamical matrix are too identical.
However, when applying the InterpolateMesh with lo_to_splitting = True, we get imaginary frequencies
for the rhombohedral dynamical matrix, while we get the correct result for the cubic.

It seems that lo_to_splitting has bug for non high symmetries structures. This needs to be checked.
Maybe it would be good to check it against the julia module.