More efficient handling of overlap matrices

[agustinsalcedo]

The overlap matrices are large and take up a lot of RAM and disk space. Two main modifications related to this have been implemented in this branch:

• Now salted.cp2k.density_fitting saves only the lower triangle of the overlap matrices, using the new save_overlap function, taking advantage of the fact that they are symmetric. When these matrices are read by salted.validation, salted.hessian_matrix and salted_minimize_loss, the full matrix is reconstructed by the load_overlap function. Note that load_overlap ensures backward compatibility, it can distinguish between lower-triangle matrices and full matrices and read them accordingly.
• The line del over, psivec added to salted.hessian_matrix prevents the script from storing two copies of the overlap matrices at the same time in RAM.

Other minor changes are:

• More timers to make the verbose output of salted.hessian_matrix more detailed.
• Re-formatting of the code using Ruff.