The Parallelism Tradeoff: Limitations of Log-Precision Transformers