We pass a lot of state around the various functions that comprise

cache_memoize() and cache_memoize_iter(), and we have a lot of key

building that we duplicate everywhere. Some upcoming work will be making

a lot of this even more complex, so in preparation for that, this change

cleans up much of this state and logic.
There's now an internal _CacheContext class, which stores the state

being used for a caching operation. This includes:

The cache connection
The base key name passed to cache_memoize()/cache_memoize_iter()
the full key name generated through make_cache_key()
The timeout
Whether to compress large data

This simplifies a lot of function signatures.
It then provides functions for generating suitable keys (wrapping

make_cache_key(), subkeys (used for storing chunks), loading cached

data, and storing cached data. Right now, these are mostly very simple

wrappers, but they'll soon be extended for additional checks and

functionality, hence the consolidation of this logic.
There's also a performance improvement/race condition fix here. Both

cache_memoize() and cache_memoize_iter() previously checked if a

value was in cache, and then pulled it out. This requires two hits to

cache instead of one, and it's possible (though unlikely) that the value

could disappear in the meantime. In practice, we'd gracefully handle the

failure, but the two cache hits weren't necessary.
We now pull the value out, defaulting the result to a _NO_RESULTS

sentinel if not found. We then make a determination based on this.
Aside from the performance improvement, there are no changes to how

the cache is utilized. This does make extensive improvements to our

documentation, and improves some logging (adding a missing "Cache miss"

for cache_memoize() to go along with that in cache_memoize_iter()),

and logging full cache keys).

All Djblets and Review Board unit tests pass.