Fixes #58

Specifically check for a match that is a repeat of the previous match
distance. If one is found, skip looking for a match in the hash table.
(This check is done one byte past the current position, as described at
https://cbloomrants.blogspot.com/2017/09/some-learnings-from-zstd.html)

Rather than searching for a match at every byte while parsing,
it now does a search first, using an algorithm inspired by M4.
This speeds it up significantly, both because there are fewer
search steps to do, and because there are fewer matches for the
parser to sort thorough. Compression improves on shorter chain
lengths, and decreses on higher ones.

Instead of trying to carry the previous match distance through
we record the length of the run of literals in length,
and literals are now distinguished by distance == 0.