Ideal Abstractions for Decision-Focused Learning