Project: Explode locations figure export

[tnagorra]

Changes

Benchmarked on the dataset with 186,321 figures, comparing the new
explode-by-locations export against the existing regular figure export.

figure export	wall-clock	peak RSS	queries	file size
regular (existing)	7.6 min	~334 MB	2	~168 MB
explode (new)	8.3 min	~334 MB	89	~194 MB

RSS = Resident Set Size (the amount of a process's memory that is in RAM)

Takeaways

• Explode runs ~9% slower than the existing export
• Memory is flat and bounded (~334 MB RSS, ~148 MB Python), independent of row
  count. Both the read and write paths stream to disk
• Query count is bounded 2 (regular) vs 89 (explode = 1 base query +
  87 batched per-chunk side queries). Query count scales with chunks, not figures.

Optimizations

• Dropped the unused locations ArrayAgg on explode
• Replaced the .filter(geo_locations__isnull=False).distinct() filter with an
  Exists() semi-join

Together these cut the explode-specific overhead by ~44% (from +72 s / 15.6%
to +43 s / 9.6% over the regular export)

Known limitation

A full unfiltered export (~8.3 min) runs close to the 10-min Celery task time_limit. Follow-ups:

• raise the export time_limit
• adopt xlsxwriter for faster writes
• validate if we reach 1,048,576-row Excel limitation

This PR doesn't introduce any:

• temporary files, auto-generated files or secret keys
• n+1 queries
• flake8 issues
• print
• typos
• unwanted comments

This PR contains valid:

• tests
• permission checks (tests here too)
• translations