Add option to use chardet for encoding detection

The tradeoff is it's much, much, much slower. In my tests, circa 10
times slower than without chardet. But it always use the right encoding.

Maybe the right thing to do is only a fallback to chardet since most of
source code is in ascii/utf-8/iso8859-1. This will be left undecided
until 1.2 comes out.