fix(server): Refresh projects based on internal timestamps

[jan-auer]

The last_fetch timestamp in project configs is set when it is written into the cache in Sentry. Since this timestamp may be arbitrarily old, this will force Relays to constantly refresh all project caches. While this is not such a big deal internally, since those requests go to a cache, it creates too much load for external Relays. Also, these time stamps are subject to clock drift and should not be trusted.

As such, the last_fetch timestamp has no useful semantics. Instead, the instant of loading the ProjectState should be used to determine the last fetch.

Since fetching projects is now a fast operation (both internally and externally), the overcomplicated logic for creating jitter is no longer appropriate. Instead, a simple time-based backoff for refreshes is absolutely sufficient.

Reverts #277

[jan-auer]

Example of what happened:

• A project state is written into the cache with last_fetch = 29 Mai 2020 17:21:38
• Relay fetches this state at 29 Mai 2020 17:42:10, but with last_fetch = 29 Mai 2020 17:21:38
• The project state is immediately soft-outdated and refreshed.

[untitaker]

Would not necessarily attempt removing the jitter in the same deploy but this makes perfect sense

[jan-auer]

Verified behavior manually for external Relays, today. We can do a staged rollout on Tuesday with @tonyo.