Add hierarchical asTuple

[bernhardmgruber]

Rename asTuple into asFlatTuple and add a new function asTuple to VirtualDatum which reflects the datum domains hierarchy.

[bernhardmgruber]

This PR just allowed us to write e.g. the nbody kernel as:

template <typename VirtualParticleI, typename VirtualParticleJ>
LLAMA_FN_HOST_ACC_INLINE void pPInteraction(VirtualParticleI&& pi, VirtualParticleJ pj)
{
#ifdef USE_STRUCTURED_BINDING
    auto [piPos, piVel, _] = pi.asTuple();
    auto [piPosX, piPosY, piPosZ] = piPos;
    auto [piVelX, piVelY, piVelZ] = piVel;

    auto [pjPos, _2, pjMass] = pj.asTuple();
    auto [pjPosX, pjPosY, pjPosZ] = pjPos;

    auto distX = piPosX - pjPosX;
    auto distY = piPosY - pjPosY;
    auto distZ = piPosZ - pjPosZ;
    distX *= distX;
    distY *= distY;
    distZ *= distZ;
    const FP distSqr = EPS2 + distX + distY + distZ;
    const FP distSixth = distSqr * distSqr * distSqr;
    const FP invDistCube = 1.0f / std::sqrt(distSixth);
    const FP sts = pjMass * invDistCube * TIMESTEP;
    piVelX += distX * sts;
    piVelY += distY * sts;
    piVelZ += distZ * sts;
#else
    auto dist = pi(tag::Pos{}) - pj(tag::Pos{});
    dist *= dist;
    const FP distSqr = EPS2 + dist(tag::X{}) + dist(tag::Y{}) + dist(tag::Z{});
    const FP distSixth = distSqr * distSqr * distSqr;
    const FP invDistCube = 1.0f / std::sqrt(distSixth);
    const FP sts = pj(tag::Mass{}) * invDistCube * TIMESTEP;
    pi(tag::Vel{}) += dist * sts;
#endif
}

Both versions compile down to exactly the same assembly: https://godbolt.org/z/r7K1hM

This alone might not look like much, but it proofs that the VirtualDatum could be replaced entirely by a std::tuple of memory locations. And this opens up some possibly interesting API designs.