event cache: unload the pinned event cache at the appropriate time

The new `PinnedEventCache` introduced recently is spawned lazily. It will spawn a task in the background as well, that will be correctly aborted when the parent container struct is dropped. However, there's nothing signaling, at the moment, that all pinned `Timeline`s are dead, and that the `PinnedEventCache` could be unloaded (i.e. the task could be aborted), resulting in unnecessary background processing, after an app has switched the view away from the pinned events timeline.

- We could likely reuse and generalize the mechanism introduced for the unthreaded room timeline (née `RoomEventCacheSubscriber`, which name would have to be generalized).
- Alternatively, we _could_ also just count pinned timelines as observers of the main room timeline — which make sense, since a pinned event timeline requires the unthreaded timeline to be alive and updated over time. Then, when the `RoomEventCacheSubscriber` drops, we could also unload the `PinnedEventCache`.
- (other million dollars idea goes here)

Start by tracing PinnedEventCache, pinned Timeline lifetimes, and the existing RoomEventCacheSubscriber mechanism. Compare the proposed observer-counting and generalized-subscriber approaches; the work is done when the pinned cache task is unloaded after all pinned timelines are gone, without stopping while one remains.

Written by the indexing model from the issue text.