ENH: new function union1d

[OmarManzoor]

Fixes #491

• Adds the union1d function.

[lucascolley]

Relevant upstream context from data-apis/array-api#25 @kgryte :

NumPy includes additional set functions (e.g., in1d, intersect1d, isin, setdiff1d, union1d, setxor1d); however, these functions are not widely implemented by other analyzed array libraries ( see here) and, thus, were not included in this initial specification. Should additional set functions be necessary, they can be proposed in follow-up proposals.

Given @paddyroddy's analysis in gh-491, it looks like absence from PyTorch might still be a blocker for inclusion in the standard? If so, probably fine to proceed with this PR.

[lucascolley]

thanks @OmarManzoor, looks good!

[lucascolley]

@hameerabbasi is there an existing discussion about sparse returning NumPy arrays for functions with naturally dense return values? Blosc does a similar thing with decompressed arrays, and there was a brief discussion at data-apis/array-api#968 (reply in thread) about whether this could be more aligned with the standard.

@OmarManzoor in the meantime, would wrapping the xp.unique_values call in xp.asarray fix the tests for sparse?