Current approach with backward optimal parsing should be left for comparison, but too slow for general use. And it needs to separate data into blocks, which makes compression worse on large files.

Should rewrite packer to use forward parsing like lzma/tornado.