Rewrite Markov implementation with backtracking optimization

[wander]

Current Markov implementation walks the prob tree each time for each candidate name (print_pwd function), and emits the word. A potentially major optimization is to walk the tree and emit words (call sha1 function) while keeping the backtracking state.

Also, consider a clean rewrite (we changed a hack, which mapped 0x00 to beginning of word and also mapped to a character in data structure, but this may have introduced a bug).