Add opt-in watch auto-reconnect helper with checkpoint resume

[rawkode]

Problem

src/client/watch.rs documents that watch streams do not reconnect automatically.

Impact

Transient network/server failures require manual reconnect logic in every consumer, increasing duplication and risk of dropped change events.

Proposed change

• Provide an opt-in watch helper/wrapper that:
  • reconnects on transient failures with bounded backoff,
  • resumes from the last checkpoint token,
  • surfaces non-retryable failures immediately.
• Keep current behavior as default for users who do not opt in.
• Add tests covering reconnect and checkpoint resume flow.

Acceptance criteria

• New opt-in API exists for automatic reconnect + resume.
• Reconnect uses backoff and resumes with the most recent checkpoint token.
• Non-retryable errors are surfaced without infinite retry.
• Tests validate reconnect behavior and no-duplication/no-gap resume expectations.