A new cache_memoize_iter function has been added that allows the
data generation function to be a generator. Unlike with cache_memoize,
the results won't need to be gathered up into a list before being handed
down to the caller, and instead, cache_memoize_iter will simply
store the cached data as items are yielded.

When data is already cached, it still functions as a generator, as
efficiently as possible. All the chunk data is still loaded and
optionally decompressed up-front, as before, since it's measurably
faster to do it then rather than on-demand. The unpickling, however, is
on-demand. We only unpickle when we're ready to parse out the next item
from cached generator.

All data passed to cache_memoize_iter is assumed to be "large
data," unlike with cache_memoize, where that is optional and up to the
caller.

Since that function now exclusively handles large data, cache_memoize
now uses it for all large data, passing in a single-item list with the
data to cache.

There are also some optimizations in here for making better use of
StringIO instead of combining/splitting byte strings.

Djblets and Review Board unit tests pass.
Browsed around a bunch of review requests and diffs, looking to see if anything

was broken. Everything worked fine. Verified in a different browser that the

data was properly cached in memory.