in particular load() and setupConnection()

Track the time before processing a request

More slow request debugging

The db writes are still instant, but the notifications are now delayed
by up to 1 second, in order to not swamp the pubsub channels.

- Allowing unlimited concurrency on OnE2EEData causes huge spikes in DB conns
  when device lists change.
- Using a high, bounded amount of concurrency ensure we don't breach DB conn limits.

With unit tests.

Add WorkerPool and use it for OnE2EEData

…q params

Because the proxy services changes to req params preferentially to live
data, if the client constantly changes the window (e.g due to spidering)
then it can accidentally stop the delivery of live events to the client
until the spidering process is complete. To help address this, we now
process live updates _even if_ we have some data to send to the client.
This is bounded in size to prevent the inverse happening: constantly
seeing new live events which starves changes to req params. This should
hopefully strike the right balance.

With regression test.

…expired connections

…t sent by that conn

…as it's more useful this way

…ount

otherwise we only see the error message if we're using sentry.

Report a metric for the size of gappy state blocks

Fix #192: ignore unseen old events