Back off GitHub server errors before exhausting job retries

[gulkaran]

GitHub server errors currently exhaust BackgroundJob's default five executions in roughly 12 seconds, so longer outages spill into the queue adapter's retries and repeatedly report errors. Use Active Job's built-in polynomial backoff with ten total attempts: about 4.3 hours plus jitter before re-raising.

The shared policy covers deployment creation, deployment statuses, and check-run refreshes. Exhausted retries still raise to the queue adapter and existing error reporter; rate-limit and authentication handling retain their existing behavior.

Tests exercise all ten attempts through serialized jobs, recovery of a failed deployment status without recreating published records, check-run backoff, and failure propagation from CreateOnGithubJob.

Validation: reproduced the short-delay regression before the fix; 1,254 tests / 3,575 assertions pass with SQLite and Ruby 3.2.0 (bundle exec rake test); repository-wide bundle exec rubocop passes (328 files).

Operational tradeoff: persistent GitHub 5xx failures take longer to reach the error reporter. Production adoption requires releasing the engine, updating Shipit's dependency, and deploying Shipit.

Related: https://github.com/shop/issues/issues/94251