Skip constrained ase.atoms in forward

[PicoCentauri]

Fixes #206

Particles in an ase.Atoms object that are part of a FixAtoms constraint are now
excluded from selected_atoms during model inference. avoiding redundant per-atom
output computation for atoms whose forces will be zeroed by the optimizer anyway.

Contributor (creator of pull-request) checklist

• Tests updated (for new features and bugfixes)?
• Documentation updated (for new features)?
• Issue referenced (for PRs that solve an issue)?

Reviewer checklist

• CHANGELOG updated with public API or any other important changes?

📚 Download documentation for this pull-request

[ceriottm]

I'm a bit confused/concerned here. Even if atoms are constrained, their per-atom energy will contribute to the forces on non-constrained atoms, so we can't really drop them from the calculation

[PicoCentauri]

We don't remove them from the calculations. Fixed atoms are still neighbors and contribute.

[ceriottm]

We don't remove them from the calculations. Fixed atoms are still neighbors and contribute.

But they don't have their own V_i, that contributes to forces on their neighbors

[Luthaf]

True, thanks for the catch @ceriottm! I think we both had a mental model with pair potentials, and did not consider higher body-order 😅

@PicoCentauri The problem is that the force on atom j f_j is d V / dr_j, and V = sum_i V_i. So in general the force on atom j arises from gradients of the energy of all atoms which have j as a neighbor, not from the energy on atom j.

We could ignore the fixed atoms/skip them in selected_atoms when computing forces directly through non_conservative_force, but at this point if might become more of a footgun/added complexity in the code for very minimal gains.