Reduce use of DataLayouts internals

[charleskawczynski]

This is basically a peel off from #1929, and extends #1932. This PR:

• Reduces the use of parent(data) and especially size(parent(data))
• eliminates get_n_items, and instead only integers are passed (when needed)
• Uses UniversalSize to leverage static parameters more
• Introduces DataLayouts.array_size and DataLayouts.farray_size, which return static sizes of the datalayouts (array_size uses a field dimension of 1, which is often helpful)

This does increase some inference failures because I've passed some additional variables through to gpu kernels using Val, which should result in statically known offsets, potentially improving performance of some kernels (likely by no more than 30%).

[charleskawczynski]

I'll bump the allocation limits, but it looks like CI is otherwise passing.

[Sbozzolo]

Look good to me, just a question

[charleskawczynski]

just a question

?

[Sbozzolo]

just a question

?

I thought I left a question in the review, but maybe I didn't click "submit". Upon re writing the question, I found the answer myself :)