Course teams occasionally upload very large files as their course
image. Before this commit, those images would be used directly in
the student's dashboard, sometimes leading to MBs worth of image
data on that page. With this commit, we now auto-generate small
and large thumbnails of configurable size. The Student Dashboard
and Course About pages will make use of this new functionality
(CourseOverview.image_urls), but the behavior of
CourseOverview.course_image_url will not change.

Note that the thumbnails are still created in the contentstore,
and sit alongside their originals.

What's included:

1. Multiple sizes, currently starting with "raw", "small", and
   "large". This falls back to the current behavior automatically in
   the case where thumbnails don't exist or this feature has been
   disabled in configuration.

2. Django admin based configuration for image sizes and whether
   to enable the functionality at all. Note that to regenerate
   images, you'd need to wipe the CourseOverviewImageSet model
   rows -- it doesn't do that automatically. This is partly because
   it's a very rare operation, and partly because I'm not entirely
   sure what the longer term invalidation strategy should be in a
   world where we might potentially have multiple themes. The
   flexible configuration was intended to allow better customization
   and theming.

3. The Course About pages also use the new thumbnail functionality,
   as an example of "large". This is in addition to the "small"
   used on the student dashboard.

Things I'm punting on for now (followup PRs welcome!):

1. Bringing the thumbnails to course discovery. A quick attempt
   to do so showed that it wasn't getting properly invalidated
   and updated when publishes happen (so the old image still showed
   up). It probably has something to do with when we do the
   re-indexing because it stores this data in elasticsearch, but
   I'm not going to chase it down right now.

2. Center-cropping. While this is a nice-to-have feature, the
   behavior in this PR is no worse than what already exists in
   master in terms of image distortion (letting the browser handle
   it).

3. Automated invalidation of the images when a new config is
   created.