feat: support tensor properties in flat batches

[Ramlaoui]

Summary

• support stacked tensor ndarrays in add_arrays_batch for molecule and atom custom properties
• flat-batch tensor sections when shapes are compatible, with clear errors for ragged shapes
• reject list/tuple tensor inputs while preserving list[str] molecule properties

Validation

• cargo fmt --all --check
• cargo check -p atompack -p atompack-py
• uv run --no-sync --extra dev --locked --with "maturin>=1.4,<2.0" maturin develop
• uv run --no-sync --extra dev --locked pytest tests/test_database.py -q
• cargo clippy --workspace --all-targets -- -D warnings
• uv run --no-sync --extra dev --locked ruff check tests/test_database.py
• git diff --check

[Ramlaoui]

Performance check for the tensor property stack, measured against detached worktrees at origin/main and origin/feat/tensor-property-batch-flat with release builds and ATOMPACK_PERF_COLOR=never.

Canonical smoke targets found and run:

• make perf-smoke-py -> atompack-py/tests/test_throughput_smoke.py
• make perf-smoke-rust -> atompack/tests/throughput_smoke.rs

Median throughput over 4 samples each:

Path	main	feature	delta
Python write add_arrays_batch	648,502 mol/s	625,127 mol/s	-3.6%
Python write add_molecules	377,789 mol/s	344,842 mol/s	-8.7%
Python sequential get_molecule	1,119,850 mol/s	1,089,962 mol/s	-2.7%
Python shuffled get_molecule	1,115,842 mol/s	1,096,318 mol/s	-1.7%
Python batch get_molecules	1,117,152 mol/s	1,150,652 mol/s	+3.0%
Python flat get_molecules_flat	5,706,899 mol/s	5,605,078 mol/s	-1.8%
Rust write	428,378 mol/s	438,570 mol/s	+2.4%
Rust sequential read	932,678 mol/s	944,008 mol/s	+1.2%
Rust random read	896,412 mol/s	934,798 mol/s	+4.3%
Rust shuffled batch read	1,503,248 mol/s	1,569,769 mol/s	+4.4%

The flat/batched read paths, which are the sensitive paths for tensor concatenation, do not show a slowdown in this smoke. The Python object write path is the only short-sample dip, but it still passes the existing smoke threshold by a wide margin. I also ran a feature-branch-only ad hoc check for explicit set_property(..., scope="atom") object writes: molecule-scope vector median ~495k mol/s, atom-scope vector median ~484k mol/s over 3 samples.

[Ramlaoui]

Update after the add_molecules performance diagnosis in the core PR:

The earlier Python object-write dip came from Python binding clone overhead, not the batch/flat tensor concatenation path. Core PR #36 now includes 792b1bb, which passes borrowed owned molecules to Rust instead of cloning every PyMolecule during Database.add_molecules.

This branch was rebased onto the updated core branch and force-pushed. Post-rebase validation:

• uv run --no-sync --extra dev --locked pytest tests/test_database.py -q -> 60 passed
• ATOMPACK_PERF_COLOR=never make perf-smoke-py -> passed

Latest smoke numbers on this branch:

• write add_arrays_batch: ~679,706 mol/s
• write add_molecules: ~585,616 mol/s
• batch get_molecules: ~1,142,901 mol/s
• flat get_molecules_flat: ~4,942,872 mol/s

Batch and flat read paths remain comfortably above the smoke thresholds, and the object-write path no longer shows the previous regression shape.