feat: Add a retry state for long-running components

[jsonbailey]

Summary

Stacked on #349. The FDv1 polling data source no longer stops for good after an HTTP error such as a 401.

• Polling waits through RetryState: one interval after a success or a normal failure, and the extended backoff (5 minutes up to 1 hour) after an unexpected failure. The next poll is timed from when the previous one completes.
• A 200 with a payload the SDK cannot use counts as a failure and clears the ETag, so the retry fetches fresh data. A 304 after a failure sets the status back to valid.
• No HTTP error sets DataSourceState.shutdown any more. The status docs are updated.
• Contract test service: declares retry-conformance-fdv1-polling, and applies polling.pollIntervalMs with no minimum, through a hook that works only when asserts are on.