feat: Retry data sources after unexpected HTTP errors instead of stopping

[beekld]

Summary

Android equivalent of ios-client-sdk#533 (SDK-2802). The streaming and polling data sources stopped permanently on 401, 403, and most other 4xx statuses, and a 401 shut down the whole client. They now retry those failures with a longer backoff than other errors: 5 minutes, doubling to 1 hour. The client stays uninitialized and keeps trying.

Implementation notes:

• RetryState is the backoff state machine shared by all the data sources.
• The SDK owns every reconnect. Extending fix: Stop the FDv1 stream on the stream thread after an unrecoverable HTTP error #402, the connection error handler now ends the stream on every failure, so the library never retries. fix: Stop the FDv1 stream on the stream thread after an unrecoverable HTTP error #402's two no-reconnect tests are inverted.
• Polling schedules each poll after the previous one completes, so a failed poll can wait longer than the interval.

Notes for reviewers:

• In FDv2 the orchestrator holds the long backoff, so fallback stays immediate: a synchronizer reports TERMINAL_ERROR on a 401, FDv2DataSource moves to the next one at once, and the failed slot returns after a backoff instead of permanently. If every slot is backing off, the data source waits rather than reporting OFF.
• Recovery to a higher-priority synchronizer waits until that synchronizer's backoff has ended.
• The SDK's data sources no longer call DataSourceUpdateSink.shutDown(). It stays in the interface for custom data sources.
• Polling failures now set isRetryable() from the status code rather than always true.
• Unlike iOS, setOnline() starts a fresh backoff, because Android builds a new data source on every state change.

Contract tests declare retry-conformance-fdv1-streaming and retry-conformance-fdv1-polling, so the harness skips its legacy "do not retry" client tests.

Testing

• RetryStateTest covers the backoff algorithm; SourceManagerTest covers slot backoff.
• The data source and synchronizer tests that asserted a permanent stop now assert a retry.