Rust acceleration via PyO3/maturin

[craftsangjae]

Summary

• Accelerate core operations (Gray code partitioning, scatter-add, empty partition filling) in Rust via PyO3/maturin
• Switch build system from hatchling to maturin with cross-platform wheel builds in CI
• Graceful fallback to pure Python when the Rust extension is unavailable

Benchmark (ARM64, Python 3.9, NumPy 2.0)

Single document encoding (128 vectors, dim=128):

Config	Rust	Python	Speedup
R=20, k=5, d_proj=8	0.69 ms	11.51 ms	16.7x
R=20, k=5, d_proj=16	0.74 ms	12.00 ms	16.2x

Batch document encoding (100 docs, ~128 vectors each):

Config	Rust	Python	Speedup
R=20, k=5, d_proj=8	17.89 ms	32.07 ms	1.8x
R=20, k=5, d_proj=16	23.42 ms	57.66 ms	2.5x

Changes

• New: `Cargo.toml`, `src/{lib,gray_code,partition,scatter,fill_empty}.rs` — Rust kernel implementations
• New: `muvera/_rust_kernels.pyi` — type stubs for the Rust module
• New: `tests/test_rust_equivalence.py` — 13 numerical equivalence tests (Rust vs Python)
• Modified: `pyproject.toml` — switch to maturin build backend
• Modified: `muvera/{init,helper,muvera}.py` — call Rust kernels with graceful fallback
• Modified: `.github/workflows/{test,publish}.yml` — add Rust toolchain + maturin-action for wheel builds
• Modified: `README.md` — add Rust acceleration benchmark results
• Modified: `examples/basic_usage.py` — use `list[np.ndarray]` input format

Test plan

• All 70 existing tests pass unchanged
• 13 new Rust/Python numerical equivalence tests pass
• `ruff check .` and `ruff format --check .` pass
• `examples/basic_usage.py` runs successfully
• Benchmark measurements completed