Fix charges output and total_charge handling for QET in PESCalculator and JAX path

[YiqingChen524]

When the DGL backend was removed, the only calculator with charge plumbing went with it. For charge-predicting potentials (QET):

• get_charges() raised PropertyNotImplementedError on both PESCalculator and JAXPESCalculator: "charges" was missing from implemented_properties and never stored in results, even though the torch Potential returns them and the JAX QEq solve computes them.
• total_charge was silently ignored everywhere: the torch calculator never passed it (LinearQeq treats None as zero) and the JAX conversion baked it as a static 0.0, so charged cells ran as neutral with no error. atoms.set_initial_charges(), which the DGL calculator consumed, was also ignored.

Changes:

• PESCalculator: "charges" in implemented_properties, results populated when potential.calc_charge is set, total_charge/ext_pot kwargs restored with the atoms.get_initial_charges() fallback. Relaxer and MolecularDynamics inherit this automatically.
• make_potential_fn(): public (E, forces, stress) 3-tuple contract unchanged by default. with_charges=True (QET only) returns a callable with a trailing total_charge argument that also returns per-atom charges; total_charge is traced, so varying it triggers no recompilation.
• qet_energy_and_charges() returns (energy, charges) and accepts a traced total_charge; qet_energy() stays a thin wrapper.
• JAXPESCalculator mirrors PESCalculator: "charges" exposed, total_charge kwarg with the initial-charges fallback.