Haystack's never been good at dealing with changes to the configured

search backend, but more recent versions have become more difficult,

with more and more objects being initialized and stored on import. This

means that our prior attempts to clear caches when changing backends

just wasn't practical anymore.
This change implements the same approach we use for cache backends. We

now have a single main backend we assign to Haystack, which then

forwards on requests to the configured backend. This gives us a lot more

control and is more future-proof.
The one thing that can't be addressed here is a change they made to

SearchView. They construct a SearchQuerySet at class definition

time, which then pulls the BaseSearchQuery from the backend and stores

it. The only workaround for this is pretty straight-forward: we

construct a new SearchQuerySet at HTTP dispatch time.
With this comes a few additional fixes, primarily in our unit tests. We

sometimes hit random breakages in search tests, due to Whoosh being used

in asynchronous mode when writing to the disk. This sometimes resulted

in newly-indexed data failing to be read ("fixed" by sleeping for .1

second) and sometimes resulted in a prior test clobbering a successive

one (fixed by giving each run its own index directory within a temp

directory).
There's also a couple packaging updates. We now depend on the latest

Haystack series prior to 3.0 (which is Python 3-only), and we depend on

elasticsearch (which is a pretty small module with a very specific

version range we require).

All unit tests pass (except some unrelated breakages I'm sorting outs

separately).
Ran the suite several times and didn't hit any more random breakages,

though I'm not ruling out the possibility.
Tested searching on both Whoosh and Elasticsearch. Switched back-and-forth

between an Elasticsearch configuration with a populated index and a Whoosh

configuration with an empty index. I was able to dynamically switch and use

the proper engine without restarting the web server. Confirmed the right

backend objects were in use via print statements.