Device list tracking is broken when using "full" syncs (using memory store) causing UTD's

[toger5]

This Issue was transferred over from EC:
element-hq/element-call#2907

This is not anymore a specific EC issue because EC only used the memory store as a stop gap and now uses a IndexedDB based store.
This is still an issue in the js-sdk.

Original Description:

There seems to be a bit of a problem in the way we track other users' device lists.

matrix-sdk-crypto keeps a cache of other users' device lists, and relies on a notification in the device_lists section of the /sync response to let it know when a user has updated their devices (ie, that cache is stale). When we send an encrypted message to another user, we use that cache to know who to send it to.

That's ok as long as we keep doing incremental syncs. The problem is that Element Call, when it starts up, does a full /sync (ie, it doesn't pass a since parameter); in that case, the homeserver returns an empty device_lists response.

So, if another user has logged in on a new device while we were offline, we end up with a stale cache for that user's devices, and they will receive UTD messages from us.

This seems like something that needs to be fixed in matrix-js-sdk. Probably the easiest solution is to invalidate all device list caches on restart, if there is no saved sync token.

Note See also element-hq/element-call#2907 for more details.

[poljar]

The native sliding sync implementation has a similar problem and we mark all tracked users as dirty when this happens. The relevant method for the OlmMachine can be found here: https://github.com/matrix-org/matrix-rust-sdk/blob/c8270cedb078155efdcbf3d78621c10c4fba3143/crates/matrix-sdk-crypto/src/machine/mod.rs#L2041-L2050.