Avoid recomputing the model on every property read in FormationEnergyCalculator

[u7k4rs6]

FormationEnergyCalculator.calculate() calls the wrapped calculator directly but never calls Calculator.calculate(self, ...), so the wrapper's self.atoms is never set. ASE's check_state then reports changes on every access, and every get_potential_energy / get_forces / get_stress call reruns the full model even when nothing changed.

In practice this hurts relaxations and MD a lot, since optimizers read energy and forces separately. With a call-counting EMT wrapped in FormationEnergyCalculator, a 10-step BFGS run did 44 base calculations on main vs 11 with this change.

The fix calls Calculator.calculate(self, ...) first so ASE can cache, and copies the check_state override from FAIRChemCalculator so changing atoms.info (charge/spin) still triggers a recompute.

Added test_formation_energy_calculator_caches_results. It uses a small counting calculator (no checkpoint needed) and checks that repeated reads hit the base calculator once, and that moving atoms or changing atoms.info recomputes. It fails on main and passes here. ruff (pinned 0.5.1) is clean.

[zulissimeta]

Great catch! Thanks for the PR.