Inserted auth_token headers aren't WSGI compliant

The PEP says [1]:

> "Native" strings (which are always implemented using the
> type named ``str``) ... are used for request/response
> headers and metadata

and

> Do not be confused however: even if Python's ``str`` type
> is actually Unicode "under the hood", the *content* of
> native strings must still be translatable to bytes via
> the Latin-1 encoding!

But when auth_token adds headers [2], it takes them straight from the auth_ref, which was deserialized from JSON and thus uses unicode. This may or may not be the right thing to do on Python 3 (I haven't looked to see whether the AccessInfo instances typically limit unicode values to the Latin-1 set or not), but it's definitely wrong for Python 2 where all headers should be bytes. This can cause applications to break in not-always-obvious ways [3].

I've attached a partial fix, but more thought seems required -- in particular there are places where we set some headers to None, which also seems out of spec.

[1] https://www.python.org/dev/peps/pep-3333/#a-note-on-string-types
[2] https://github.com/openstack/keystonemiddleware/blob/5.2.0/keystonemiddleware/auth_token/_request.py#L162-L168
[3] https://bugs.launchpad.net/swift/+bug/1793880