Implement lazy area weights

[schlunma]

🚀 Pull Request

Description

This PR makes iris.analysis.cartography.area_weights lazy by providing the keyword arguments compute and chunks:

def area_weights(cube, normalize=False, compute=True, chunks=None):

These defaults ensure full backwards-compatibility. chunks=None results in using the cube data's chunks for the weights array, which is a resonable default since these weights will probably be used with the cube's data anyway.

This requires https://github.com/SciTools/iris/pull/5620, so still in draft mode. Apart from that, this PR is ready.

Closes https://github.com/SciTools/iris/issues/5611

---

Consult Iris pull request check list

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 89.77%. Comparing base (7c313ff) to head (e6985e3). Report is 4 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5658   +/-   ##
=======================================
  Coverage   89.76%   89.77%           
=======================================
  Files          93       93           
  Lines       22982    22995   +13     
  Branches     5006     5011    +5     
=======================================
+ Hits        20630    20643   +13     
  Misses       1622     1622           
  Partials      730      730           

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[trexfeathers]

Thanks for the hard work @schlunma. #5620 will be reviewed by @pp-mo in the new year, so #5658 can get moving again once that is sorted.

[CLAassistant]

All committers have signed the CLA.

[schlunma]

Just had a second look on this. I think it's actually better to use cube.lazy_data().chunks instead of cube.lazy_data().chunksize to make sure that the chunk structure is exactly identical in the weights and the original data. It probably won't matter, but it feels better to use .chunks here...

[schlunma]

After having a third look on it, I think it's better to leave full control of the chunks to the user (i.e., if chunks=None, also use chunks=None in the broadcast_to_shape call instead of the cube's chunks). Otherwise, there is no way a user can specify to use "no special chunking", which might be confusing.

Sorry for all the back and forth, with this I am happy with the PR!