Investigate runtime data layout performance

[Ivorforce]

I've looked into this a bit. The relevant bit is here.

It uses a runtime bool to test for linear assignment, and constexpr for whether to use simd assignment. So far, so good.

We should still check if it actually does enter linear_assign sometimes.

[Ivorforce]

I've gotten a bit further with this. #126 was fixed for array fills. I've also forked xtensor for xtensor-stack/xtensor#2810. I think this resolves the issue for now, since I've exhausted all spots where I think contiguous_layout is used (the static equivalent of is_contiguous). I've also opened PRs for all relevant spots in xtensor.

The only thing left to do now is to keep eyes open for potential similar pitfalls in the future.