Benchmarking Generalization via In-Context Instructions on 1, 600+ Language Tasks