feat(index): batch invalidation writes for improved performance

[hicder]

Change invalidation persistence strategy from immediate disk writes to in-memory buffering with periodic flushing. This reduces I/O overhead when processing document deletions.

Key changes:

MultiSpannIndex:

• invalidate() and invalidate_batch() now only update in-memory state
• Add flush_invalidations() to persist buffered invalidations to disk
• Track pending invalidations in DashMap<u128, HashSet<u128>>

MutableSegment:

• Buffer invalidations in DashMap during invalidate() calls
• Write all pending invalidations to disk during build() via InvalidatedIdsStorage

ImmutableSegment:

• Add pending_invalidations: RwLock<HashMap<u128, HashSet<u128>>>
• Add flush_invalidations() method called during segment operations
• remove() now buffers instead of immediately persisting

Collection:

• Call flush_invalidations() on finalized segments during flush

io_uring improvements (uring_engine.rs, uring_file.rs):

• Fix buffer pinning: use Pin<Vec<u8>> instead of Pin<Box<Vec<u8>>>
• Replace std::sync::Mutex with tokio::sync::Mutex for async safety
• Add explicit Send/Sync impls for UringFile
• Fix Clippy warning: use for loop instead of while let

This design batches disk writes for deletions, reducing I/O overhead while maintaining consistency by flushing during segment build/flush operations.