consider doing fewer retries for APNS

[richvdh]

if we have problems with APNS, we retry each request at 10s and 20s intervals. Given that haproxy limits the number of concurrent requests to sygnal, this can easily mean that the haproxy<->signal connection pool gets saturated with failing APNS requests, so that very few GCM requests get a look-in.

We should consider doing fewer retries, or doing them faster, or marking APNS as down and failing the request more quickly.

Really, we need a mechanism to indicate to homeservers that a push failed (https://github.com/matrix-org/matrix-doc/issues/2534).

[richvdh]

Also: apnspushkin makes 3 attempts, and then for each attempt, aioapns makes 4 attempts (we set pool.max_connection_attempts=3, but I think there is an off-by-one error in aioapns). Is all that really needed?

[richvdh]

see #104 for a bit of background on this