feat: support tensor custom properties

[Ramlaoui]

Summary

• add tensor custom property storage tags with per-value shape metadata
• allow molecule and atom custom tensor properties through set_property(..., scope=...)
• make get_property/property_keys/has_property/delete_property work across molecule and atom scopes with unique keys
• keep tensor shape out of schema equality so same-key tensor values can vary by record until a concatenating API asks for uniform shapes

Validation

• cargo fmt --all --check
• cargo check -p atompack -p atompack-py
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• uv run --no-sync --extra dev --locked pytest tests/test_atom_molecule.py tests/test_database.py -q
• uv run --no-sync --extra dev --locked ruff format --check python
• uv run --no-sync --extra dev --locked ruff check python

[Ramlaoui]

Follow-up performance diagnosis for add_molecules:

The apparent slowdown was not the Rust storage path globally. The Rust smoke with builtin energy/forces stayed flat or faster, and an isolated Python split showed builtins-only object writes were also fine. The slowdown appeared when Python-owned molecules carried custom properties.

Root cause: the Python Database.add_molecules binding cloned every owned PyMolecule before writing. Adding tensor support increases PropertyValue from 32 bytes to 56 bytes, so even legacy scalar/vector custom properties became more expensive to clone. The canonical Python smoke prebuilds molecules outside the timed region, but the timed add_molecules call still paid this clone cost.

Fix added in 792b1bb: pass borrowed &Molecule references for owned Python molecules instead of cloning them; raw SOA view handling is unchanged.

Validation after the fix:

• cargo fmt --all --check
• cargo check -p atompack-py
• uv run --no-sync --extra dev --locked pytest tests/test_database.py tests/test_atom_molecule.py -q -> 70 passed
• ATOMPACK_PERF_COLOR=never make perf-smoke-py -> passed, write add_molecules ~611,725 mol/s

Ad hoc split on the patched branch:

• builtins-only object writes: median ~760k mol/s
• custom-property object writes: median ~622k mol/s

So the tensor feature exposed an avoidable clone cost; the branch now removes that cost instead of accepting the regression.