unicode ssl errors

[stefanfoulis]

fix for #24

This PR DOES NOT WORK YET. I'm opening it to start a discussion on how we should approach this problem. My current approach seems tedious. There must be a better way.

There are a lot of subtle problems with python3 (the testsuite does not pass yet).
The testsuite passes on python2 (except for AWS4Auth_GetCanonicalHeaders_Test.test_headers_amz_example which did not pass before my changes either).

This PR might not be necessary at all if kennethreitz/requests#3573 is fixed. Although there might be more subtleties, because:

# python 2
>>> x = {u'a': 42, b'b': 42}
>>> b'a' in x
True
>>> u'a' in x
True
>>> b'b' in x
True
>>> u'b' in x
True

# python 3
>>> x = {u'a': 42, b'b': 42}
>>> b'a' in x
False
>>> u'a' in x
True
>>> b'b' in x
True
>>> u'b' in x
False

current python3 testsuite output: https://gist.github.com/stefanfoulis/1c9b46a61beba2e6423caa746006cae0