Skim loading fails running with sharrow when unused skims are included

[dhensle]

Describe the bug
When running with sharrow enabled and a clean sharrow cache, _drop_unused_names scans CSV config files for skim variable references and drops any skims not found. Later, reload_from_omx_3d reads directly from the OMX files and tries to write all matrices (including the dropped ones) into the shared memory dataset, causing a KeyError.

This only happens when:

• sharrow is enabled
• The sharrow cache is clean (no cached zarr)
• The skim data doesn't require realignment (so the load=False / reload_from_omx_3d path is taken)
• Some skims in the OMX files are not referenced in any CSV spec file

To Reproduce
Steps to reproduce the behavior:

1. run with sharrow on
2. Ensure OMX skim files contain a matrix that is not referenced in any csv file in the config directories
3. Ensure land use TAZ alignment matches the skims exactly so the load=False code path is taken

Screenshots