[DRAFT] fix: skip builtin field names when copying atoms.arrays/info

[speckhard]

Summary

_extract_ase_record (from_ase path) AND is_reserved_ase_array_key (to_ase path) both copied/emitted keys for builtin field names without filtering, causing silent duplicate state on round-trip:

• from_ase: atoms.arrays["forces"] and atoms.info["energy"] would land in custom properties alongside the calculator-derived builtin values.
• to_ase: a custom property named after a builtin (e.g. via mol.set_property("forces", X)) would land in atoms.arrays["forces"] next to the calculator-attached builtin forces.

This PR fixes both directions.

What changes

atompack-py/python/atompack/ase_bridge.py:

• _merge_properties (called for atoms.info and the info= override) now skips all _BUILTIN_FIELDS. The existing stress-as-(3,3) builtin override behavior is preserved inside the new _BUILTIN_FIELDS branch.
• The atoms.arrays loop previously skipped only {"positions", "numbers"}. Now also skips _BUILTIN_FIELDS.
• atoms.calc.results already filtered _BUILTIN_FIELDS correctly (line 186 on main); no change needed there.

atompack-py/src/molecule_helpers.rs:

• is_reserved_ase_array_key previously matched only "numbers" | "positions". Now also rejects "energy" | "forces" | "charges" | "velocities" | "cell" | "stress" | "pbc".

Out of scope (deferred)

The set_property side still allows the user to set a custom property named after a builtin (only "stress" is reserved at set_property time today). Tightening set_property to reject all builtin keys is a more invasive behavior change worth its own PR — the present defensive fix already prevents the symmetric duplicate-state on to_ase regardless of what set_property accepts.

Backward compat

• If anyone was relying on atoms.info["forces"] becoming a custom property after from_ase, that behavior is gone; use the dedicated from_ase(forces=...) kwarg instead.
• If anyone was relying on mol.set_property("forces", X) then surfacing in atoms.arrays["forces"] after to_ase, that behavior is also gone; the calculator-attached forces are now the only path.
• No on-disk format change. No public API signature change.

Test plan

Three new tests:

• test_from_ase_does_not_duplicate_builtins_into_custom_properties — hostile atoms.arrays and atoms.info with builtin-named keys; confirms builtins remain calculator-derived and no leak into custom properties.
• test_from_ase_info_override_kwarg_filters_builtins — same filter applies to the info= kwarg path of from_ase.
• test_to_ase_does_not_duplicate_builtins_in_arrays — mol.set_property("forces", X) does NOT contaminate atoms.arrays["forces"] after to_ase; legitimate non-builtin custom property still flows.

$ uv run --extra dev --locked --with "maturin>=1.4,<2.0" maturin develop --release
🛠 Installed atompack-db-0.2.1

$ uv run --extra dev --locked pytest tests/test_from_ase.py -v
============================== 15 passed in 0.47s ==============================

$ uv run --extra dev --locked pytest tests/ -q
142 passed, 1 skipped in 7.03s   # 139 baseline + 3 new tests