Answer
Orchestration: one coordinator tells each service what to do and runs the compensations when a step fails. The flow lives in one place, easy to follow, change and monitor, but the orchestrator is one more service to run. Choreography: each service reacts to events and publishes its own. No central piece and loose coupling, but the flow is spread across services, hard to see and test.
Why
Choreography suits two or three steps; longer flows usually move to an orchestrator or a workflow engine (Temporal, AWS Step Functions), which also keeps the saga's state across crashes.