ENH: make __setitem__ raise on a device mismatch

[ev-br]

closes gh-207

[ogrisel]

LGTM: I checked that the current scikit-learn test suite passes against that branch of array-api-strict.

I also checked that changing scikit-learn code to allocate a result array with xp.empty without setting the expected device= kwarg make the tests fail early, when calling __setitem__ time, as expected.

[ev-br]

data-apis/array-api#1005 is I believe the spec update to make this spec-compliant.

[ev-br]

data-apis/array-api#1005 has been merged, so this PR is unblocked. Formally, the spec update will only be effective in the next release, which, if it's release history is any guide, can be expected in early 2027.

While the remit of array-api-strict is to scrupulously follow the official, published spec to the letter, delaying this PR by at least half year feels impractical and unhelpful for the ongoing adoption effort. So let's let make a pragmatic move and land this now.
Thanks Olivier for the review.