Add retry/backoff to the CI and gate-verdict pollers

[JSONbored]

Problem: A single transient error (e.g. a brief 5xx from GitHub) kills the entire poll loop for CI status and gate verdicts — the poller only re-attempts while a conclusion is genuinely "pending," not after an error.

Area: AMS / Miner

Proposal: Add bounded retry with backoff around the poller's own HTTP calls, distinct from its existing "pending" polling logic.

Deliverables:

• Retry/backoff wrapping the poller's fetch calls, with a sane max-attempts ceiling.

Acceptance criteria:

• A simulated transient error mid-poll is retried and the poll completes successfully instead of aborting.

Resources:

• The CI-poller and gate-verdict-poller modules.

[linear-code]

JSO-3489