WIP: refactor: remove modulestore global

[ormsbee]

Let's see how much stuff breaks if we just remove the cache modulestore global.

This could introduce a performance regression if it's called many times (it looks like it takes something like 1-2 ms per invocation). If it is a problem, request-level caching may be enough to take care of it.

Holding onto global modulestore references has long been a source of memory leaks and gives us many opportunities to shoot ourselves in the foot when dealing with multiple threads and potential cross-contamination of the shared object across multiple users. This could remove a whole class of bugs if it works. Of course, it could also break things in unanticipated ways.

There are a bunch of places where we call modulestore() repeatedly in the same function instead of assigning to a variable, on the assumption that it's a singleton. We would want to clean that up. It might involve passing the modulestore as an optional argument in a bunch of places, so we stop recreating it so casually, like:

def some_fn(foo, bar, store=None):
    store = store or modulestore()
    # rest of function

The risk is performance regression from calling modulestore init functions many times.

It's also possible that we can make modulestore init even cheaper than it is today.