#123 Fix: Implement incremental checkpoint commits

MEDIUM-TERM FIX for progress preservation during crashes. Problem: Checkpoints are committed atomically with task completion. If worker crashes mid-execution, all progress is lost. Changes needed: 1. Add checkpoint_commit_mode option (atomic vs incremental) 2. For incremental mode: commit checkpoints in separate transaction 3. Add checkpoint versioning for conflict detection 4. Update DurableContext to support incremental commits 5. Ensure idempotent replay from last committed checkpoint Trade-off: Incremental commits add overhead but preserve progress. Parent: #116