Fix D3 wrapping for periodic positions

[timduignan]

Summary

• pass wrap_positions through the nvalchemiops neighbor-list helper
• make AlchemiDFTD3 request wrapping internally so direct D3 calls are invariant to unwrapped periodic coordinates
• add a regression test for integer-cell translated periodic positions

Why

AlchemiDFTD3.forward is a public wrapper boundary and can be called directly by dynamics code. It previously relied on callers to pre-wrap periodic positions, which can silently undercount D3 interactions for unwrapped coordinates.

Tests

• /Users/timothyduignan/Projects/orb-meta/SIM/libs/core/.venv/bin/python -m pytest tests/forcefield/test_d3.py -q
• /Users/timothyduignan/Projects/orb-meta/SIM/libs/core/.venv/bin/python -m ruff check orb_models/common/atoms/graph_featurization.py orb_models/forcefield/inference/d3_model.py tests/forcefield/test_d3.py
• git diff --check