DataShard: write lock conflicts may be lost on restarts without in-memory migration

[snaury]

While investigating #27123 I found that reads with an active write lock may reply while failing to persist additional conflicts. Such conflicts may then be lost and allow commits to violate serializability.