Behaviour for pluggable subscription-state backends.

Subscription checkpoints track WHERE each consumer is in a stream (or global log). They are written frequently — every ack — but are small and not latency-critical. Most storage backends can handle them fine; the behaviour exists so that a backend can choose whether to keep checkpoints in its own store or delegate to the Khepri control plane.

The default implementation (reckon_db_khepri_subscription_backend) stores checkpoints in Khepri. This is the natural fit — subscription state IS a control-plane concern, and Ra consensus on checkpoints is exactly what you want for "don't replay already-acknowledged events on leader failover."

Fast log backends (RocksDB, custom) will typically implement reckon_db_log_backend but NOT this — they defer to the Khepri implementation. Having the behaviour explicit makes the split clean.