Zebrad: add rollback tip height command

[endophysics]

Adds zebrad rollback-tip-height --height <height> to safely roll back local state by rebuilding a bounded copy through verified block replay, then atomically swapping it into place while keeping the previous state as a backup.
Added it because during sync I would sometimes get stuck on a local branch when mining was enabled, which would then force me to restart the process.

Also invalidates tip-dependent GUI wallet/Zaino caches after rollback and fixes staking/delegation replay during checkpoint copying so rebuilt state preserves accounting correctly.

[Zk-nd3r]

Swap-and-restore instead of in-place finalized mutation is the right call.

Two on the staking replay:

1. checkpoint_replay_bond_rewards / checkpoint_replay_reward_total reimplement the reward math the non-finalized path already does via finalized.bond_rewards. Two copies of consensus math will drift. Can it call the same function instead of mirroring it?
2. The active-bond set is computed separately in chain.rs and delegation.rs. If they ever disagree, the pool tally and per-bond credits desync. Worth a test that replays a staking block and checks the derived rewards match the non-finalized output.

[Zk-nd3r]

Nice, this looks like it addresses the staking replay concerns from my earlier comment. bond_rewards_for_active_bonds is now the single reward-split source, checkpoint_replay_reward_total is gone, and the finalized pool tally now comes from the same per-bond replay output. I also checked the live call sites and they all feed POS_BLOCK_REWARD_ZATS.

The one test gap I still see is checkpoint replay end-to-end: the new test covers the helper path, but not a staking block driven through both non-finalized and checkpoint/finalized replay with equal bond rewards/value pools. I think that would be worth adding.

One separate rollback edge: rollback_tip_height.rs swaps the active DB before resetting tip-dependent wallet/Zaino caches. If cache reset fails, the command returns an error after rollback, leaving stale caches beside the new tip. Maybe reset caches before the swap, or restore the old DB if cache reset fails?

Testing: git diff --check origin/dev...HEAD passes. The clean zebra-state --lib test target is currently blocked by unrelated NearTipChains compile errors outside this PR.