Move device changes off the main process

[sandhose]

The main goal of this PR is to handle device list changes onto multiple writers, off the main process, so that we can have logins happening whilst Synapse is rolling-restarting.

This is quite an intrusive change, so I would advise to review this commit by commit; I tried to keep the history as clean as possible.

There are a few things to consider:

• the device_list_key in stream tokens becomes a MultiWriterStreamToken, which has a few implications in sync and on the storage layer
• we had a split between DeviceHandler and DeviceWorkerHandler for master vs. worker process. I've kept this split, but making it rather writer vs. non-writer worker, using method overrides for doing replication calls when needed
• there are a few operations that need to happen on a single worker at a time. Instead of using cross-worker locks, for now I made them run on the first writer on the list

[MadLittleMods]

Actually adding /delete_devices I think uncovered two more bugs:

• deleting access token was only available on the main store, not worker store (fixed with 07f32aa)
• when we called notify_device_update through replication, it was missing a conversion from StrCollection (which can be a lot of things) to a proper list to be serializable. Grrr @ replication clients that don't retain proper typing… Fixed with d916e9e

-- @sandhose, #18581 (comment)

Overall, this is the type of thing I can't really review (related discussion: #18581 (comment))

I can add an approval but it's just like "sure".

Feels like some typing/lints are missing around ensuring only serializable data is passed to a ReplicationEndpoint but this is something to address outside of this PR.