fix: detect abnormal shutdown and trigger repair on next start

[guillaumemichel]

Fix suggested in #279 (comment)

Problem

A CRDT replica can be left in a silently inconsistent state after an abnormal termination (SIGKILL, OOM, panic without Close(), power loss).

Fix

Persist a single key, /<namespace>/bs, whose presence at startup means "the previous run did not close cleanly":

• Startup: if the key is present, MarkDirty the store (so repair runs) and log a warning. Then write the key unconditionally so a crash in this run is caught on the next start.
• Clean Close(): after wg.Wait() (so any worker still marking dirty during shutdown has finished), delete the key as the last step.

No changes to the walk/merge hot path, MarkDirty/IsDirty signatures, or repairDAG scope.

[gammazero]

One question. Otherwise ok to merge.