refactor(lr): migrate PME Coulomb to energy+autograd

[isayev]

Fixes #106.

Summary

Completes the #105 migration for PME: a single energy-only nvalchemiops call leaves positions, charges, and cell in the autograd graph, and every derivative (inference forces/stress, force/stress training losses, dense Hessians, HVPs) flows through the calculator's total-energy autograd.

The #106 blocker was an upstream composed-graph charge-gradient bug in nvalchemi-toolkit-ops 0.4.0. The 0.4.1 release reworked exactly the implicated machinery — the charges.grad_fn-based _has_potentially_geometry_dependent_charges heuristic (whose instability matches the issue's "charge-gradient edge sometimes None, sometimes valid" anomaly) was replaced by a deterministic _dechain_connected_input_grads pass that removes each dependent-input path exactly once. Verified with the calculator-level train-vs-eval force comparison from the issue (spiro crystal, tests/data/2000054.cif, single L40S):

nvalchemiops	PME energy-graph route: max abs dF (train vs eval)	force elements off by >1e-3
0.4.0	4.477 eV/A	100%
0.4.1	5.7e-6 eV/A	0%

Mean abs F is 3.75 eV/A on this system, so the 0.4.0 corruption averaged ~21% of the force magnitude — six orders above the ~1e-6 GPU noise floor — while total energies agreed to 1e-6 and Ewald through the identical code path was clean on both versions. With the bug fixed upstream, the issue's file-a-minimal-reproducer-upstream step is moot.

Changes

• aimnet/modules/lr.py: PME joins the Ewald energy-graph branch. Removed the legacy explicit-terms path, the _PeriodicCoulombFunction training wrapper, _periodic_coulomb_hybrid, and the fixed-charge FD Hessian/HVP helpers; ExternalDerivativeTerms drops its now-unproduced hessian field.
• Runtime guard: selecting PME on nvalchemiops < 0.4.1 raises RuntimeError with the upgrade command. The pyproject pin only applies at install time; without the guard an already-installed 0.4.0 would corrupt training forces silently.
• aimnet/calculators/calculator.py: the training_derivatives flag is now DSF-only; the PME strain-input plumbing (_external_strain_inputs) is removed; the HVP drops the PME directional-FD block (PME joins the Ewald exact-autograd float64 contract; eps is retained but unused); stale docstrings rewritten.
• aimnet/calculators/derivatives.py: external-terms Hessian merge removed (no producer remains).
• pyproject.toml: nvalchemi-toolkit-ops>=0.4.1.
• Tests: legacy-path tests replaced with energy-graph equivalents parametrized over ewald+pme — module-level autograd-vs-FD force gate, autograd-Hessian-vs-force-FD gate, energy-in-graph contract spy, and version-guard tests. The pre-existing eval-vs-train, Hessian sum-rule, and HVP-vs-dense gates now cover PME's new path unchanged.
• Docs: long_range.md and calculator.md derivative-support sections rewritten (both also predated the refactor(lr): migrate Ewald Coulomb to energy+autograd #105 Ewald migration and still claimed NotImplementedError Hessians).

Semantic change

PME dense Hessians and HVPs switch from fixed-charge finite difference to relaxed-charge autograd: they now include the d^2E/(dq.dr) charge-response coupling, matching DSF and Ewald, so vibrational frequencies / IR are comparable across all LR backends. Values shift slightly against the former FD numbers. This also retires the last upstream DeprecationWarning (deprecated direct-output flags).

Note for existing environments

Environments with 0.4.0 already installed must pip install -U nvalchemi-toolkit-ops before selecting PME; the guard raises with that instruction until then.

Review round

A final review pass over the finished diff surfaced and led to: the uv.lock regeneration (it still pinned 0.4.0 against the new >=0.4.1 floor), several stale docstrings/comments describing the removed FD machinery, an honest rephrasing of cross-backend Hessian comparability (Ewald and PME are directly comparable; DSF still differs by its shifted-force truncation near the cutoff), disclosure of the dense-PME-Hessian dtype change (float64 FD block -> force dtype, matching Ewald since #105; the HVP keeps its documented float64 return), guard hardening (pre-releases of too-old versions such as 0.4.0rc1 are now refused; only truly unparsable versions pass), and an eps DeprecationWarning on hessian_vector_product.

Testing

All in an nvalchemi-toolkit-ops 0.4.1 environment, serial on one L40S:

• Full suite: 700 passed / 4 skipped on the bulk, plus 127 (test_lr+test_pbc), 151 (test_calculator+test_hvp), and a final 210-test re-run of the two post-review-edited files — 0 failures.
• Slow create_graph=True gates run explicitly: test_ewald_pme_hessian_finite_symmetric_sumrule[pme], test_hvp_matches_dense_periodic[pme], test_ewald_pme_hessian_consistent_with_forces[pme], test_hvp_periodic_returns_float64[pme], test_eval_matches_train_forces_and_stress[pme]: pass.