Add Figshare download script and standardize fixtures to 50 rows

[doncamilom]

Summary

• scripts/setup_data.py: end-to-end Figshare download, extraction, and MIST_DATA_DIR setup
• Standardized all reaction task fixtures to 50 rows (rxn_inversion 12->50, rxn_naming 20->50, rxn_truefalse 12->50)
• demo/expand_fixtures.py: utility for expanding fixtures by cycling rows
• Updated fixture_manifest.csv with correct counts

scripts/setup_data.py

Downloads datasets.zip (2.2 GB) and optionally models.zip (5 GB) from Figshare, verifies MD5, extracts, and maps to the ${MIST_DATA_DIR} layout expected by recipes. Writes .env.local for convenience.

# Datasets only (skip 5GB model download)
python scripts/setup_data.py --data-dir ./data --skip-models

# Full setup (datasets + models)
python scripts/setup_data.py --data-dir ./data

# Then either:
source .env.local && export MIST_DATA_DIR
# Or:
export MIST_DATA_DIR=./data

Test plan

• Verify python demo/run_fixture_smoke.py passes with all fixtures
• Verify scripts/setup_data.py --skip-models --skip-verify runs (dry-run on CI without downloading 2GB)