Use copy-on-write shape and strides

[Ivorforce]

Currently, compute_case uses xt::dynamic_shape for shape and strides.
Since these often don't change for operations, it should be possible to use a CowData - like structure to hold these instead. This should reduce unnecessary allocations. For primitive shapes (like Vector2, i.e. float[2]), we can even avoid allocation altogether by using a static shared shape for construction.

I should measure the impact before endeavouring on this, but I suspect it should decrease overhead a fair bit.