Include long-range geometry in PolarMACE stress

[darthjaja6]

This fixes #1642.

The local MACE graph was already strained when stress was requested, but the long-range PolarMACE path was still using the original positions, cell, reciprocal cell and volume. That meant stress only saw the indirect response of the learned charge density and missed the explicit geometric derivative of the electrostatic energy.

This gives the long-range calculation the same affine positions and cell as the local model, then recomputes the reciprocal cell and volume from that strained cell. The original positions are still kept as the force differentiation variables.

I added an 8 Šperiodic water case to the finite-difference stress test. It failed by about 4.3e-5 eV/ų before this change and now agrees within 2e-7. The existing force case and all four periodic-box stress cases pass, and asking for stress does not change the energy or forces.

This currently pins graph_electrostatics to the commit from graph_electrostatics#5, because reciprocal-vector gradients need that small autograd fix. I’ll switch the pin back to the upstream repository once that PR is merged.

[aacostadiaz]

Thanks for this. The fix looks good to me.

Waiting for WillBaldwin0/graph_electrostatics#5 to merge so we can flip the pin back to WillBaldwin0.