ENH: linalg module, mT property, and multiple outputs from @compiled

[mtsokol]

Hi @hameerabbasi,

This PR introduces improvements that my pydata/sparse use-case requires:

• Array API compatible linalg module with the first function, vector_norm.
• Array API mT array class property.
• @compiled decorator accepting multiple outputs.

[hameerabbasi]

Suggested a few simplifications to the code. The overall logic of the PR LGTM.

[hameerabbasi]

A test for multiple-return would also be appreciated.

[mtsokol]

A test for multiple-return would also be appreciated.

Done!

[willow-ahrens]

I'm surprised that @compiled didn't already accept multiple outputs! The interface you've implemented is really intuitive. Thanks Mateusz.

[mtsokol]

@hameerabbasi Thanks for improving the nesting returns logic!
Left last comments, and maybe one last thing is a test which covers:

@compiled()
def foo_bar(a: Tensor, b: Tensor, c: int):
    c = a + b
    d = c ** 2
    return (a, c, (b, "text"), {"key1": 12, "key2": d})

[hameerabbasi]

The test caught some bugs! Thanks for the suggestions, @mtsokol.