That's already done in _populate_hosts_and_request_paths().

With multiple API versions, the class may be different. It's also
conceivable that different connection classes may eventually be mixed
and matched with different driver classes, especially in the case of
OpenStack w/ Keystone.

Version-specific subclasses will inherit from this. I expect more will
be moved into it in coming commits.

Note 1: No "Base" in the class name because it is useable as-is.
Note 2: The OpenStack base classes I'm writing will be JSON-based. Not
wanting to rewrite the (soon-to-be-cut?) XML-based 1.0 driver, that will
instead inherit and override for XML where necessary, such as happens
here.

I noticed this problem frequently: a Connection's request() instantiates
and response, then sets the responses connection attr to itself;
however, many of the Responses have __init__s that call their parse_body
or parse_error which refers to its own connection object to get details
for error output. Fail. Have to pass the connection to the constructor
to be available early enough.

Courtesy of PyDev.

HTTP spec states that the Host header must contain the port number if
it's non-standard (see
http://www.w3.org/Protocols/rfc2616/rfc2616-sec14.html#sec14.23). The
incorrect override that was present here was breaking the managerment
urls in OpenStack auth responses. httplib does the right thing if we let
it.

Also, the "connection =" removal was because it's shadowed by a
following assignment.

Auth 1.1 is not necessary. Auth 1.1 servers accept 1.0 authentication,
which gives just as much data (i.e. the same management API URLs).
Note: I'll fix the tests later.

The Connections now also allow a tenant_id (which appears to be the new
name for project_id, so I'll go with it) to be passed in and honored
(this is necessary for Alpha customers).

Not all Auth versions need the /auth (e.g. Alpha doesn't). Better to use
the full URL as the setting so they can be treated the same.

At least for the OS compute classes, that is. At some point, I hope
someone fixes the other OS service drivers and moves this baseward.

Since most OS stuff will be JSON-based, I put a shared encode_data as
far baseward as I dared without risking other stuff (someone should look
at the other services, though). The 1.0 driver is XML-based, so I had to
override it there. As long as I was doing that, I cleaned up all of the
places where each and every method had been doing the encoding itself.

All basic ops, most of the extended ones. I'll almost certainly add a
few more.

This effectively makes the OpenStack driver appear as a single driver
that supports all versions, specified during instantiation (except it's
much easier to maintain).

Plus a little bit o' PEP8.