Rethinking Variable-Length Encoding: Exploiting Bit Sparsity for Parallel Decoding in LLM Accelerators | Synapse