Refactor windowed arrays to use a numpy backed DataArray cache

[wyatt-fluidnumerics]

Description

This improvement is largely outlined in #2843. The goal is to remove the current codes dependence on a np.stack call in WindowedArray.isel() which causes a large in memory copy. This implementation replaces the existing dictionary cache with a numpy backed DataArray, which allows for direct indexing and removes the need to reconstruct a DataArray from the dictionary cache on every call to .isel(). This results in significant performance improvement for large datasets (as much 20x or more depending on size and advection scheme).

Additionally, these changes mean that in theory WindowedArray now has all the needed functionality to support non-synchronous clocks. In practice however, this is likely not a very practical use case as for large datasets holding even a few time levels in the cache can lead to OOM errors on many machines.

Checklist

• Closes Windowed Array Improvements/Optimization #2843
• This PR targets the correct branch (main for normal development, v3-support for v3 support)

AI Disclosure

• This PR contains AI-generated content.
  • I have tested any AI-generated content in my PR.
  • I take responsibility for any AI-generated content in my PR.
  • Describe how you used it (e.g., by pasting your prompt): Claude code was used to write code to benchmark this change against the previous version and create the plot in Windowed Array Improvements/Optimization #2843. I thoroughly reviewed the benchmarking code. It was also used to write parts of the new implementation, though significant changes and rewrites to those components were made by myself.

[erikvansebille]

Looks good!