Fix/forest type document statistic

[tsubik]

Speed up document statistics counting

generate_for_country_and_day counted in Ruby: the join-alias columns
(fmus.forest_type, required_operator_documents.required_operator_document_group_id)
went through method_missing at ~33µs per read vs ~0.31µs for a real attribute,
the triple loop rescanned every document once per combination, and
previous_stat ran one query per combination.

Now Postgres does the counting — a single GROUPING SETS query per day covers
every country and every rollup — and previous stats load in one DISTINCT ON.

Measured over a full rebuild:

• unoptimized, with the forest-type fix applied: ~208s per day across all six
  scopes (~60h for 1,049 days)
• after this PR: 0.43s per day — 497 days rebuilt in 3.6 min

Worth noting the two are linked: the old code was only fast because the broken
forest-type comparison short-circuited on the first check. Fixing it makes far
more comparisons succeed, which made the unoptimized version dramatically
slower — so this optimization is what keeps the fixed code usable.

Output is unchanged. Rebuilt 2023-03-08 → today under the old and new
implementations through an identical loop and compared with EXCEPT ALL both
ways: 42,538 rows each, zero differences — including the set of rows whose date
was moved forward.

Also included: forest-type slices in statistics were always zero (same
String/Integer mismatch, broke when Fmu.forest_type became an enum in March
2023), sync:operator_documents_stats swept with delete_old: false so slices
stopped summing to their rollups, and a one-off fix:erroneous_document_deletion
task for the two documents this bug deleted.

🤖 Generated with Claude Code