Use obstore in data processing workflows

[spencerkclark]

This follows the approach we have taken in the ERA5 data processing workflow to use obstore when reading/writing from/to zarr in the cloud. It is expected this should provide a meaningful performance improvement, particularly when reading from stores with small inner chunks, namely the stats and time coarsening parts of our workflow.

Test workflows on the same 10-year dataset:

• Without obstore: compute-fme-dataset-ensemble-fzl9k
• With obstore: compute-fme-dataset-ensemble-cjp5g

Timing results:

	Dataset computation	Stats computation
Without obstore	48m	34m
With obstore	38m	35m

Surprisingly we do not see any meaningful change in the stats computation time, though we do see a faster dataset computation step.

Changes:

• Adds a get_zarr_store function in compute_dataset.py and get_stats.py and uses it in compute_stats.py, get_stats.py, and time_coarsen.py.
• Updates the processing image to include obstore. The new image is tagged v2026.03.0 and we have updated the argo workflow to use the image in all steps.