Parser module uses hand-written parser (much faster). Get rid of parsec

[Paczesiowa]

this makes prepare function much faster (and much less garbagy). I've tested the new implementation with the old one using a few runs of 10^6 testcases generated with quickcheck

[zenzike]

I'm a little nervous about merging this one in: it may be much faster (by how much?), but it doesn't seem nearly as maintainable to me. If the concern is that parsec is too slow, have you looked at using attoparsec to solve this?