Add specification for computing a tensor contraction (linalg: tensordot)

[kgryte]

This PR

• specifies the interface for performing a tensor contraction.
• is derived from comparing signatures across array libraries.

Notes

• JAX supports providing a precision keyword argument to control matrix-multiplication precision. It is alone in doing so among compared array libraries.

• TF allows providing a (2,K) shaped array for the axes keyword argument. NumPy et al support array-like values or lists when specifying multiple axes. This proposal follows NumPy et al in only allowing int and Sequence.

• TF returns an array whose data type is that of x1. This proposal requires that the result follow type promotion rules.

[kgryte]

This PR has been open for some time without comment and discussed/approved during meetings. Will merge and we can submit follow-up PRs should any issues/concerns arise.