chore(data): remove third-party raw data from git — fetch instead

[setupelz]

Summary

• Untrack 14 third-party raw data files (PRIMAP, CMIP7, WDI, UNU-WIDER Gini/WDI, UN/OWID population, LULUCF, GCB bunkers, AR6 Gidden scenarios); they stay on disk and are fetched via fair-shares fetch-data.
• Add exact-path .gitignore entries (no directory wildcards) so CITATION.md files stay tracked in each directory.
• data/regions/*.csv, data/rcbs/*, and data/scenarios/iamc_example/*.xlsx remain tracked (own compilations / cleared example).

Lint & Test Results

• Lint: ruff check . — 116 pre-existing findings, unchanged from main (unrelated to this change, out of scope)
• Tests: pytest tests/ -v — 742 passed, 11 deselected
• pytest tests/unit/test_data_registry.py -q — 58 passed (committed-data-tree test class skips cleanly with data untracked)

Test Plan

• git status shows 14 staged deletions + modified .gitignore, nothing else
• Files verified still present on disk after git rm --cached
• git ls-files data/ | grep -i citation — 10 CITATION.md files still tracked
• git ls-files data/ — no remaining .nc/.xlsx/.csv/.feather/.zip except data/regions/*.csv and data/scenarios/iamc_example/*.xlsx
• Full test suite green

[setupelz]

CI is red on all three Python versions (3.11/3.12/3.13) — 10 failures, not a flake:

• 6 in tests/integration/test_python_api_reproduces_notebook.py (co2-ffi-EPC, co2-ffi-ECPC, all-ghg-EPC-free-rider, all-ghg-EPC-co2-debit, all-ghg-PCA-capability, all-ghg-ex-ECPC)
• 4 in tests/unit/test_rcb_generator_config.py (test_rcb_pathways_default_generator, test_rcb_pathways_explicit_generator, test_non_rcb_pathways_without_generator, test_integration_config_and_source_id)

All fail identically: ConfigurationError: Emissions data file not found: data/emissions/primap-202503/Guetschow_et_al_2025-PRIMAP-hist_v2.6.1_final_no_rounding_13-Mar-2025.nc.

Root cause: .github/workflows/test.yml states its own now-invalidated assumption — "Auto-fetch is disabled in tests (tests/conftest.py), so CI never downloads data." That design assumed the raw data files were present in the checkout because they were git-tracked. This PR untracks them, so a fresh CI checkout has no data and these 10 tests (which exercise real emissions data, not just registry/config plumbing) fail at config-validation time.

tests/unit/test_data_registry.py (the class this operation was verified against) does skip cleanly — 58/58 pass. The 10 failures are outside that class and have no existing skip-on-missing-data marker (checked: only pyam-availability skips exist elsewhere in the suite).

Local pytest tests/ -q shows 742 passed because the files are still present on disk locally — that check can't catch this, since the change's entire effect is on a fresh checkout with no local data cache.

This needs a maintainer decision before merge, since either option is a design change, not a git operation:

1. Add a fetch-data step to CI (can't fully replicate local — GCB is licence-gated/manual per this PR's own commit message, so CI would still need a strategy for that source or would need to skip GCB-dependent tests specifically).
2. Mark these 10 tests as requiring real data and skip them in CI (drops notebook-reproduction integration coverage from CI, which may be load-bearing e.g. for the JOSS paper).

Not merging until this is resolved — main is currently green (last run: success) and this PR would turn it red.