Deal with recurring Sentry reports of temporary server shutdowns

Sentry has tons of reports from RPC clients that are caused by their servers being temporarily down (typically, a postgresql database).

For example: https://sentry.softwareheritage.org/share/issue/42e620d6e6194d9e985fe2e3e525f080/

Instead, we should make servers and clients cooperate so they know this is a temporary "expected failure" (ie. similar to HTTP 503), and make clients retry for longer.

---

Migrated from T4427 (view on Phabricator)