Align casting methods with Python behaviour

[honno]

Resolves #486 (adding x.__complex__()).

Per interest in having the builtin casting methods (__bool__/__int__/__float__/__complex__/__index__) support arrays of a different dtype kind, this PR also changes the language to clarify that adopters can support such casting. EDIT: now aligning casting methods with Python behaviour generally

(I opted against requiring cross-kind casting in this PR—see #486 (comment)) See #486 (comment)

[honno]

Per what we decided upon in our last array meeting (see #486 (comment)), I've updated this PR with my take on how to we align the spec's casting methods with Python's behaviour.

I have some thorough testing for how Python's casting works in https://gist.github.com/honno/98371048b4d5b9fc4a003377a714e827, which I'll have to tidy up (will be useful anyway for the test suite).

As @seberg noted, NumPy seems to 100% follow Python casting behaviour for its arrays (if we're not including a small divergence in behaviour for NumPy scalars). Only notable thing maybe is NumPy not aligning with Python's __index__ (not actually casting I suppose), in that it raises TypeError for bool arrays, where Python bool objects just returns itself.

Not 100% on language used, happy to hear what folk think. Will sleep on it anyway.

[honno]

Just realised I should probably special case non-finite values for __bool__.

[honno]

Think this is ready @seberg @asmeurer? With the follow up of specifying <bool-array>.__index__() behaviour as erroneous (context #497 (comment)).

[honno]

Note previews are broken because of some work I was doing in my fork for #505, but you can change the URLs to lead with .../0/build/draft/index.html like so still.

[kgryte]

I believe all discussion points have been resolved. I'll go ahead and merge. Any further changes can be addressed in follow-up PRs. Thanks, all!