Fixing Hackable Benchmarks for Vision-Language Compositionality