Fix alignment for Kokkos::complex

[crtrott]

Reported at hackathon, in order to get proper 8 byte and 16 byte loads for complex<float> and complex<double> the user derived his own classes from CUDA's float2 and double2.

But it is enough to tell the type that it is aligned: alignas ( 2*sizeof(RealType))

[crtrott]

Needs to fix both pair and complex

[dhollman]

Don't you mean 2*alignof(T)? Especially for pair, using sizeof could be a huge cost.

[nliber]

And that won't work for pair, as it takes two different types

[dhollman]

I actually think we don't even want to go there with pair. When T and U are different, alignof do you use? If T is first, you will always get alignof(pair<T, U>) >= alignof(T). But if alignof(U) > alignof(T), what should the default behavior be?

[dhollman]

@nliber jinx

[crtrott]

We can partially specialize for pair<T,T>.

[dhollman]

I don't know if we really want to support use cases that abuse pair in this way, but if we do, we should go all in on it:

template <class T, class U>
struct alignas((alignof(T) > alignof(U)) ? 2 * alignof(T) : 2 * alignof(U)) pair;

That should do the trick, and should support whatever use cases we would have wanted to support for the pair<T, T> case

[nliber]

There look to be a fair number of uses of pair<T, T> inside Kokkos, so we probably should do this (the full general one @dhollman wrote above).

[crtrott]

I am fine with that.

[ndellingwood]

Related PR #2259