lzw: parse/process input in blocks

This avoids creating an HBytes for each and every code word. Instead, the
code words are collected into blocks behind each clear code and translated
together into a single HBytes per block.