WIP: Improve perf of applying zarr agglomerate mappings

[fm3]

URL of deployed dev instance (used for testing):

• https://___.webknossos.xyz

Steps to test:

• abc

TODOs:

• [x] compare against hdf5 on local laptop (hdf5 with cumsum about 40% faster)
• [x] profile applying zarr agglomerates (10% bucket scanner, 20% build map including chunk loading on cache misses, 70% apply)
• [ ] reimplement apply in native bucket scanner
• [ ] will the cache misses dominate then? does increasing chunk size make sense? or cumsum cache after all?

Issues:

• fixes #

---

(Please delete unneeded items, merge only when none are left open)

• [ ] Added changelog entry (create a $PR_NUMBER.md file in unreleased_changes or use ./tools/create-changelog-entry.py)
• [ ] Added migration guide entry if applicable (edit the same file as for the changelog)
• [ ] Updated documentation if applicable
• [ ] Adapted wk-libs python client if relevant API parts change
• [ ] Removed dev-only changes like prints and application.conf edits
• [ ] Considered common edge cases
• [ ] Needs datastore update after deployment