This is a situation where we have a Python unicode string which doesn't
consist entirely of genuine Unicode characters -- some of the codepoints
in the string are surrogate codepoints, which occur in a UTF-16 encoding
of a string and were also repurposed in PEP 383 for losslessly encoding
arbitrary mostly-UTF-8 bytestrings (like Unix filenames) in Python
strings.  Currently, on Python 3, we cause a UnicodeEncodeError if we
try to encode such a string as JSON.

It's not 100% obvious what the right thing to do here is -- this
situation seems like it must reflect a bug somewhere else in the
program or its environment.  But

 * one way we can get such a string is by loading a JSON document
   (perhaps an invalid JSON document? anyway, we load it without error):

   >>> ujson.dumps(ujson.loads('"\\udcff"'))
   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
   UnicodeEncodeError: 'utf-8' codec can't encode character '\udcff' in position 0: surrogates not allowed

 * we already pass these strings through without complaint on Python 2;

 * as the included test shows, passing these through matches the
   behavior of the stdlib's `json` module.

So it seems best to pass them through.

Fixes ultrajson#156.