In general, we have cyclic import issues all over the place, this is one
easy fix and will help out in later commits.

Note that this maintains backwards compat due to how the the functions
are imported into `plotly.py`.

This is also some setup for a larger refact. Since v1 and v2 requests
handle errors differently, it’s easier if we simplify the api into our
errors so that the requests can go from `response` —> `python error` as
they see fit.

This was driving me nuts. We basically manually handle creating and
validating *each* api response inside each calling function. Even worse,
we *sometimes* raise a `PlotlyRequestError` and *sometimes* just bubble
up the `requests.exceptions.HTTPError` ;__;.

This does the following:

* Define an `api.v1` module that only includes `clientresp` (the only old
api method we still *need* to cling to)
* Define an `api.v2` module that includes all the new functionality of
our v2 api.
* Both `v1` and `v2` raise the same `PlotlyRequestError`, so that users
only need to catch *one* exception class in scripts.

Note that the `apigetfile` did some weird things to convert old-style
plotlyjs figures (e.g. `type: ‘histogramx’`) to new-style versions.

I wanted to ween us off the old api, so this makes the change from
`/apigetfile` —> `/v2/plots/[fid]/content?inline_data=true`.

The `_swap*` functionality was copied from `plotly/streambed` code,
directly from the backend’s implementation of `apigetfile`.

There is a circular import issue that needed to be fixed here as well.

This is allowed when setting the credentials *file*, however, trying to
set this inside a session used to fail.

The `requests` package kindly manages 2/3 compat for json for us, might
as well be consistent and use the same tool!

As a side note, I’d like to move away from `six` and depend directly on
`requests.compat`. I believe it has everything we need and then we can
ditch the `six` dep and know that we’re always in sync with whatever
requests is doing (which is really what we care about).