Completes the remaining nunique item in #100.

Summary

• move the public nunique wrapper and backend dispatch into _delegation.py
• delegate eager NumPy and CuPy arrays to native unique(..., equal_nan=False)
• delegate eager PyTorch arrays to native unique, while retaining the generic fallback for devices such as meta that do not support data-dependent shapes
• preserve the JAX static-size implementation and the lazy/array-agnostic fallback for Dask, strict, sparse, and wrapped namespaces
• add a backend dispatch test that fails if supported native namespaces reach the generic implementation

The equal_nan=False argument preserves the existing Array API behavior where separate NaNs count as separate unique values.

Validation

• pytest -q -p no:cacheprovider: 4347 passed, 8984 skipped, 18 xfailed, 2 xpassed
• pytest -q -p no:cacheprovider tests/test_funcs.py::TestNUnique: 38 passed, 72 skipped in the lean environment
• expanded focused backend run before slimming the environment: 75 passed, 34 skipped, 1 expected Dask xfail (NumPy, array-api-strict, Dask, sparse, PyTorch CPU/meta, and JAX CPU available)
• ruff check on changed files
• ruff format --check on changed files
• numpydoc lint on changed implementation modules
• validate-pyproject pyproject.toml
• public export consistency test

CuPy could not be installed on this macOS host; its current native unique signature was checked to support equal_nan and CI covers the backend matrix.