Contact the server's .well-known API before anything to get
available login flows instead of blindly assuming it will be
m.login.password, and to get the server's real URL instead of
requiring users to remember that e.g. it's "chat.privacytools.io"
and not just "privacytools.io" despite user IDs making it look like so.
The server field will also now remember the last accepted URL.
While a user file write operation #1 was queued in the _write_loop,
it was possible meanwhile for another function to read the file
(the current old version on disk), add some data to it,
add submit the write (operation #2).
The modifications from operatio #1 were then completly lost at best.
Compression with Pillow can take long, especially with large
clipboard PNG images.
Doing this in a separate process prevents the async event loop from
getting blocked, and allows multiple compression operations to run in
parallel.
Making the window narrow enough to switch to "mobile mode", then big
again would trigger a signal in SettingsView that switched focus to it,
even when it wasn't visible.
It's possible to get a MatrixForbidden error when trying to retrieve a
offline room member's presence, if that member actually left the room
but the server hasn't yet sent the update.
This required us to set the media downloaded local path on events
entirely from python instead of simply lazy-fetching them when needed
from QML, due to pyotherside's async nature and files that must be open
in a certain order.
A healthy client must not only have a sync task running, but also have
finished its initial sync and have been succesful in its last sync
attempt.
The previous weak requirement caused this issue:
Have two clients, one on a failing server, and try to fetch a media,
thumbnail or profile: since these functions don't need auth, any
healthy client is picked to do the request, but the previous
requirement made it possible to pick the failing client.
