Migrate NequIP integration to nequip >=0.7 (NumPy 2 / Python 3.13 support)

[k-yoshimi]

Background

abICS currently pins nequip >=0.5.6, <0.7 (pyproject.toml) because the abICS
nequip solver/trainer (abics/applications/latgas_abinitio_interface/nequip.py,
nequip_trainer.py) target the pre-0.7 nequip API
(nequip-train / nequip-deploy, nequip.data.AtomicData, nequip.utils.Config,
torch.jit.load("deployed.pth")).

This series uses numpy.in1d, removed in NumPy 2, so it requires numpy<2.
Because Python 3.13 has no numpy 1.x wheels, the NequIP path cannot run on
Python 3.13, and the ActiveLearnNequip CI job is therefore skipped on 3.13
(see #81).

Goal

Migrate the abICS NequIP integration to a numpy-2 / Python-3.13 compatible
nequip (>=0.7, currently up to 0.18.0; requires-python >=3.10, numpy
unconstrained). nequip 0.7.0 is "a major backwards-incompatible update with
breaking changes throughout the code", so this is a real migration, not a
version bump.

Scope (likely)

• Rewrite nequip.py (solver) and nequip_trainer.py for the new nequip API
  (new CLI nequip-compile / nequip-package instead of nequip-deploy,
  new config format, new data structures).
• Update nequip-allegro (the pinned 0.3.0 is tied to old nequip; newer
  Allegro has a different package/API).
• Update tests/integration/active_learn_nequip inputs and install_nequip.sh,
  and re-enable the ActiveLearnNequip CI job on Python 3.13.
• Update the JA/EN docs and lift the <0.7 cap in pyproject.toml.

References

• nequip CHANGELOG: 0.7.0 breaking rewrite; 0.12.0 drops Python 3.9.
• Related: Fix Python 3.13 test failures (RXMC NumPy 2 bug; document/skip NequIP) #81 (skips/documents the current limitation).