fix: handle cache lock conflicts and clean up orphaned native state

[runchen0919]

Summary

• Add a three-stage retry strategy for redb database lock conflicts (try → retry after 500ms → delete & recreate) to prevent initialization failures when a previous process didn't release the lock
• Detect and shut down orphaned native handles during nativeInitialize when the same cache_dir is reused (e.g., ClassLoader replacement that skipped nativeShutdown)
• Call BazelBridge.shutdown() on bundle stop instead of keeping the bridge alive, relying on disk cache for classpath recovery on next restart

Test plan

• Verify normal cache open path still works (no lock conflict)
• Simulate lock conflict by opening the redb file from another process, confirm retry and recreation works
• Confirm orphaned state is cleaned up when reinitializing with the same cache_dir
• Verify bundle stop properly shuts down the native bridge
• Run cargo test --workspace and mvn test

🤖 Generated with Claude Code