When loading information for an extension, we attempt to look up one of

the metadata files, normalize the string content, and parse it. On

Python 2, this worked fine no matter what type of string format we got

from pkg_resources, because both binary and Unicode strings happily

let you run decode(). This isn't the case on Python 3. We always have

Unicode strings there, which don't support decode().
We weren't really set up to test this too well, because in the unit

tests, we were overriding too much of the metadata fetching logic,

making it hard to really test these differences.
This change updates the metadata handling to check the string type

before deciding whether we need to decode anything, and to override the

lowest-level function we can to return string content, allowing more of

pkg_resources's parsing code to run. To do this, we no longer create

our own stub for a Distribution, but instead use Distribution as-is

and give it a fake Provider, which is the backend class that looks up

metadata from a package for the distribution to use.
There's a couple of other small things that were also incorrect in this

process. We had to deal with the result type from

get_metadata_lines(), which is a generator on Python 3. Things were

also more sensitive when it came to including the admin site's URLs, and

so a test was updated to do this correctly.

Relevant unit tests pass on Python 2.7, 3.4, 3.5, and 3.6.