Add Cholesky function specification

[kgryte]

This PR

• specifies the interface for performing Cholesky decomposition.
• is derived from comparing signatures across array libraries.

Notes

• NumPy (along with CuPy, JAX, MXNet, TF) does not allow returning either the lower- or upper-triangular Cholesky factor. However, SciPy, Torch, and Dask do support returning either. The ability to return either factor is common outside of the PyData ecosystem (MATLAB, LAPACK). Accordingly, the decision was made to include an upper keyword to support returning the upper-triangular Cholesky factor.

• Following Torch, MXNet, TF, NumPy, and JAX, this proposal allows for providing a stack of square matrices.

[leofang]

It is straightforward for CuPy to support the upper keyword, as internally we delegate to cuSOLVER/rocSOLVER which supports both modes: https://docs.nvidia.com/cuda/cusolver/index.html#cuSolverDN-lt-t-gt-potrf

[kgryte]

Thanks, @leofang, for the review. This is ready for merge...

[lucascolley]

What was the motivation for choosing the upper=False default here?

[kgryte]

@lucascolley See the API comparison linked to in the OP: https://github.com/data-apis/array-api-comparison/blob/0459e5dd51fa38df8bf24363f4fa5895ac5c2929/signatures/linalg/cholesky.md

TL;DR: the kwarg was not universally supported. Among array libraries, PyTorch had upper=False, so we followed PyTorch.

[lucascolley]

Thanks @kgryte . cc @mdhaber for awareness