Previously we would always write out the data to a file, and then
call the parser on it.

Most of this is refactoring to split the file-handling logic (ie reading
from a FILE*) from stream-handling logic (encoding, counting line numbers,
etc).  One tricky point is handling encoding markers inside strings-to-
parse; we (and CPython) handle the file case by creating a Python file
object to iterate over it, and CPython handles the string case with a special
function to pre-process and decode the string.  This commit takes the
approach of creating a cStringIO file-like-object and running it through
the same decoding logic as the file case.