Fix issue #6 and switch to pixi + rattler build system

[lohedges]

This PR closes #6 by fixing issues related to the definition of physical atoms and the removal of improper torsion terms. A unit test is added for a problem perturbation from the TYK2 test set, which previously gave invalid modifications.

The PR also switches the build system to pixi and rattler-build.

[lohedges]

Summary of additional tweaks/fixes applied on this branch:

1. Independent per-state physical neighbours — removes the cross-state ghost filter that was downgrading triple→dual
   junctions in PTP1B.
2. Angle optimisation fallback — preserves original theta0 with k_soft when minimisation fails, rather than using
   partially-converged geometries (fixes CAT-13i series).
3. Ring bridge stiffening off by default — avoids strain and free energy bias for ring bridges where the ring geometry
   already constrains the ghost (fixes CAT-13f-CAT-13o).
4. sp2 bridge stiffening off by default — avoids ~30° strain for sp2 non-ring bridges (potentially fixes EJM49-EJM31 and
   similar cases).
5. Rotamer hybridisation fix — checks hybridisation at the correct end state to avoid false positives from transmuting atoms
   (fixes nitrile cases).
6. Anchor selection — scores physical neighbours to avoid anchoring through transmuting or bridge atoms (fixes PTP1B anchor
   choice).