[BUG]: Remove the reader refresh data race in FUSE read-after-write [LiAuto]

[yangcx000]

Describe the bug

Concurrent reads on one read/write handle can simultaneously observe a changed writer version, flush and reopen the reader, and call RawPtr<FuseReader>::replace without synchronization. Another task may concurrently call path or read through the same pointer.

This is Rust undefined behavior and can cause double free, use-after-free, heap corruption, or crashes.

To Reproduce

1. Open one file handle for reading and writing.
2. Perform a write that advances write_ver.
3. Issue concurrent reads on the same handle with multiple FUSE tasks.
4. Allow both reads to enter the reopen branch.

Expected behavior

Reader state must use safe ownership. Version check, writer flush, reader reopen, installation, and version update must form a synchronized per-handle state transition.

OS Version (please complete the following information):

• Linux
• Curvine branch: current main
• Component: curvine-fuse/src/fs/state/backend_handle.rs

Additional context

Add deterministic single-flight reopen tests, refresh-failure tests, concurrent read/write stress tests, and read-path performance measurements.

Dependency: the stable release/handle cleanup work should land first to reduce lifecycle conflicts.