The PageCache implementation wasn't properly dealing with timelines that
could return statuses in non-chronological order.
For example, if you bookmark a recent status, then go back in the
timeline and bookmark an older status; the bookmarks timeline is ordered
by the time of the bookmark event, not the creation time of the status
that was bookmarked.
If a sufficiently old status was bookmarked so it straddled a page
boundary you could have a situation where the range of status IDs in two
different cached pages overlapped.
E.g., this log extract:
```
0: k: 110912736679636090, prev: 3521487, next: 3057175, size: 40, range: 112219564107059218..110912736679636090
1: k: 111651744569170291, prev: 3049659, next: 2710596, size: 40, range: 111926741634665808..111651744569170291
```
The range of IDs in page 0 overlaps with the range of IDs in page 1.
The previous `PageCache` assumed this couldn't happen, and broke in
various interesting ways when it did.
E.g., you can't find the page that contains a given status by looking
for the largest key less than the needle's status id. Given the pages
above looking for ID 112219564107059218 (first status in page 0) would
suggest page 1 as having the greatest key less than that ID. This
manifested as the correct page briefly appearing in the UI (page 0),
then being completely replaced with page 1.
Rewrite PageCache to fix this. The previous implementation used a single
`TreeMap` assuming items were always sorted by ID. The new code keeps an
unordered map from status IDs to the page that contains that status, and
a separate `LinkedList` that contains the pages in order they're
provided by the API. This decouples the ordering of pages in the cache
with the overall ordering of items within the pages.
b0d63e4243