fix universal newline parsing (which the codecs module doesn't support), use the fast 'io' module for file reading in Py2.6 and later

scanned stream is unicode, so use unicode strings for comparisons to avoid redundant recoding in Py2

--HG--
extra : rebase_source : d9445f29b6d417ef02c53f7f9b490843de81501c

__Pyx_PyObject_IsTrue now has only a single branch, even at -O0.
(The bitwise | is intentional.) This is only a 2.5% or so speedup
in my microbenchmarks, but it's something.