Do some work on MSC4147

[andybalaam]

MSC4147

The work that was planned for it is described at https://github.com/element-hq/crypto-internal/issues/369

[uhoreg]

Looking at my own Synapse database, and running the following query:

select length(message_json) from device_inbox;

there seems to be two main clusters of message lengths: one around 1400, and one around 2400. So it looks like adding the device keys increases the message length by a factor of about 1.7, or by about 1000 bytes.

I'm not sure what other statistics we need to gather. We could look at matrix.org performance graphs between end-of-June (when the feature was initially implemented in the Rust SDK) and now, but I'd expect them to not be very useful as there would be a lot of factor affecting performance. So aside from saying that "it hasn't been a disaster on matrix.org", I'm not sure what else we can say.

We could add some logging? Maybe log how many to-device messages and the total size of the to-device messages are sent per /sync response, along with the total size of the /sync response, to give an idea of how much it's inflating the /sync response size on average?

I'd expect it to not have a huge impact on the speed of the /sync response, since it isn't making any queries any more complicated. It's increasing the data size, but I don't think that the data transfer is the slow part of the /sync.

[andybalaam]

The rest of this is covered by https://github.com/element-hq/crypto-internal/issues/369