Cache SplitMongo course structures in memcached.

This is primarily to reduce load on MongoDB, where we've lately
had performance problems that we suspect are caused by very
large course structures being evicted from MongoDB's cache. This
may potentially give us a path to better performance as well,
but that's not the goal of this commit.

Surprisingly, LZ4 seemed to actually run more slowly than zlib
for this. Possibly because of some overhead in the Python
bindings? GZip was also surprisingly slow given that it uses
zlib underneath (something like 5x slower).

Use separate cache backend for caching structures.

Abstract out course structure cache.

add datadog metrics for compressed course structure sizes

Since we're using a different cache background, we don't need to have a cache prefix

Use dummy cache backend for tests.

Fallback to default cache if course_structure_cache doesn't exist.