Improving Generalization on the ProcGen Benchmark with Simple Architectural Changes and Scale