feat(client): make consumer checkpoint lookup bounded and observable

[Migorithm]

Changes

• Limit rebalance checkpoint lookups to ranges being started and their transitive split/merge predecessors.
• Use the consumer-group generation as a shared snapshot boundary across independently routed range lookups.
• Require each range leader to observe the durable generation seal before returning a checkpoint.
• Serialize checkpoint lookup with commits from the recovering consumer.
• Keep range actors stopped when checkpoint lookup fails.
• Surface background lookup failures through Consumer::next_record().
• Return lookup failures directly from explicit stale-commit recovery.
• Document the rebalance checkpoint boundary and failure behavior.

Why

Rebalances previously requested checkpoints for every known topic range. Individual range requests also lacked a common observed boundary, and lookup failures were only logged before an internal retry.

The new flow makes lookup cost proportional to the ranges being provisioned, gives the result set one logical generation boundary, and makes failures visible without treating unavailable state as a missing checkpoint.