Bifurcated Attention: Accelerating Massively Parallel Decoding with Shared Prefixes in LLMs