feat: scenario families, wind bias, and train/test split

[aryamurray]

Summary

• Scenario families: ignition patterns (center/edge/corner/multi_cluster), severity levels (low/med/high), asset layouts (A: single cluster, B: two clusters)
• Wind bias: directional fire spread via (wx, wy) vector -- downwind spreads faster, upwind slower
• Train/test split: 9 frozen train families, 6 held-out families per impl-plan section 6
• Observation: extended with severity one-hot (3) + wind vector (2), total 636

Test plan

• All 24 scenario combinations (4 ignition x 3 severity x 2 layout) reset and step correctly
• Randomized sampling produces variety across resets
• Wind bias creates directional spread asymmetry
• Ruff lint clean