Add matrix_power specification

[kgryte]

This PR

• specifies an interface for raising a square matrix to an integer power.
• is derived from comparing signatures across array libraries.

Notes

• The output array dtype for NumPy depends on both the input array and whether n is negative. If negative, then the result is floating-point. If non-negative, the result has the same dtype as the input. For CuPy, only float32 and float64 is allowed. In general, we recommend restricting linalg functions to floating-point dtypes. That practice is followed here.

• Following Torch and maintaining consistency with other linalg interfaces operating in square matrices, a stack of square matrices is accepted.

[rgommers]

For CuPy, only float32 and float64 is allowed. In general, we restrict linalg functions to floating-point dtypes. That practice is followed here.

Only specifying float32/64 sounds right, however the phrasing is a bit off. I'd rather add a note saying that the behaviour is only specified for these dtypes, and that for other dtypes implementation may differ - they will either cast or raise for some or all values of n.

[kgryte]

Re: dtypes. We'd need to add that note more generally at the top of the linalg spec, as this is generally true for most linalg specs, given their current support being "limited" to floating-point dtypes. The restriction to floating-point dtypes was discussed a couple times during meetings and found its support there.

[kgryte]

@rgommers re: dtypes. We could change the language from must to should. As in, we could recommend floating-point input, but allow implementations to accept, e.g., integer dtypes.

In this particular case, I would, however, advocate for always returning a floating-point dtype.

[kgryte]

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

[kgryte]

For reference, inclusion of this API in the standard was discussed here.