WALNUT: A Benchmark on Semi-weakly Supervised Learning for Natural Language Understanding