Convert `ke::flattened_hierarchy` into a struct of arrays

[kammce]

Now, technically the Chinese remainder theorem approach will be faster but it has a size cost for the additional hierarchy information needed. So in those cases, we may still want to optimize the flatten RTTI as well as the dynamic cast algorithm. I'm not sure if this will actually improve performance but its something to look into. My guess is that this will NOT make a difference in terms of optimization because skipping two words is as simple as adding 8 to the iterator/pointer. But there may be some optimization due to this that I'm not considering.