Avoid expensive casts to numpy arrays under the hood

[erikvansebille]

When DataArray.data is called, the return is an Array-API compatible array (i.e., Numpy, but also dask). Since the ChunkCachedArray only implements a subset of the Array API, this would mean that direct calls to .data would trigger a very expensive cast to an Array API compatible array (i.e., a numpy array).

I think these casts should be avoided (in the codebase, but also importantly in interpolators) when running with CachedChunkArray's. I'm not sure what the fix would be at first glance

Originally posted by @VeckoTheGecko in #2908 (comment)

[VeckoTheGecko]

Including below my comment from #2912

---

Though there is the problem that users writing interpolators should never really be using da.data when using ChunkedArrays (since it falls back to Dask, which is less performant, or (before this PR) uses Numpy, which causes eager computation of results.

Implementing get_duck_array at all could result in users writing interpolators that have really bad performance.

I'm thinking maybe the solution is just to do a raise NotImplementedError here. This would mean that users can't inspect the data using a da.data, but I think thats acceptable (users won't be inspecting this anyway).

[erikvansebille]

This is now fixed in #2916