Add specification for computing the qr factorization

[kgryte]

This PR

• specifies the interface for computing the qr factorization.
• is derived from comparing signatures across array libraries.

Notes

• Following Torch, MXNet, and TF, this proposal allows for providing a stack of square matrices. NumPy, Dask, and CuPy do not currently support providing stacks.

• NumPy supports multiple factorization "modes": raw, reduced, complete, and r modes (along with several legacy modes). Dask does not support any modes. TF only supports reduced and complete modes. MXNet only supports reduced mode. Torch linalg.qr (latest master) does not support raw mode.

  This proposal only includes support for reduced and complete modes.

• NumPy et al provide support for factorization modes via a mode keyword. TF supports modes via a full_matrices keyword, similar to SVD. This proposal uses a mode keyword for possible future support of additional modes.

• While Torch linalg.qr supports r mode, it still returns an empty array for q (due to easier JIT). As this proposal does not include r mode support, a namedtuple containing q and r is always returned.

• NumPy, CuPy (?), and JAX support returning h and tau arrays containing the Householder reflectors and associated scaling factors, respectively. TF and Torch do not. This proposal does not support returning h and tau.

[rgommers]

Thanks @kgryte. I agree with all the choices made.

[kgryte]

Thanks, @leofang and @rgommers, for the reviews! This PR is ready for merge...