Specify output array data types

[kgryte]

This PR

• specifies the output array data types for each functional API.

Notes

• the current draft includes one TODO, which concerns the output data type when returning counts from unique.

[rgommers]

This looks like a good idea to me.

[asmeurer]

This looks good to me too.

For the "either float32 or float64" cases, should we require any kind of consistency, e.g., an array library always has a "default float dtype" that is used in these cases? Also, should it at least require the same float dtype when the argument(s) themselves promote to a float dtype (e.g., float32 + float32 should always be float32 and float64 + float64 should always be float64)?

[rgommers]

For the "either float32 or float64" cases, should we require any kind of consistency, e.g., an array library always has a "default float dtype" that is used in these cases?

Yes, I believe I suggested that somewhere as well. That would also cut down on repeated content.

Also, should it at least require the same float dtype when the argument(s) themselves promote to a float dtype (e.g., float32 + float32 should always be float32 and float64 + float64 should always be float64)?

I think that's already covered by type promotion rules, no need to spell that out.

[kgryte]

@rgommers This PR (and OP) have been updated. Should be ready for formal review.

[rgommers]

Thanks @kgryte, this is pretty comprehensive, overall looks good. A couple of small issues.

[kgryte]

Addressed feedback.

[rgommers]

LGTM now, merging. Thanks @kgryte!