Make UKF implementation work for dynamic dims

[sphw]

The original goal of this PR was to make the UKF work with Python, but it turns out the changes needed were way more than I expected. So I'm splitting the PR into this one, which makes the necessary changes to nox and roci-adcs, and a second one for the Python integration.

For nox the major push has been to relax type bounds that only worked with ConstDim. For the most part this means instead of getting the dimension with D1::DIM you instead call D1::array_shape(&arr.buf). In addition I made the following changes:

• Separate array.rs into mod.rs and a repr.rs file
• Delete the XlaDim trait, and replace all its uses
• Rename Dim in ArrayDim to Shape to disambiguate the type level dimension and the potentially dynamic shape
• Add a new map function to Tensor which maps just along the top-level dim
• Add a new function iterate over the rows of a Tensor
• Make concat_many work with iterators instead of fixed length arrays

In roci-adcs the Nox changes were used to replace all the uses of Const<N> with generic dimensions. I have kept the Const api as a wrapper, because that is safer for Rust users. There is a loosely typed internal API that the Python wrapper will use.