feat(memory): add cross-process locking and atomic writes for sqlite

[yunasora]

Fixes #3022

Implementation Details & Context

During the implementation, I aligned the changes with the existing codebase architecture and made a few adjustments based on the repository's current state:

1. Database Configuration & Connection Pool:

   • The requested parameters (PRAGMA journal_mode=WAL, busy_timeout=5000, and SetMaxOpenConns(1)) are already centrally configured in sqliteutil/sqlite.go (which corresponds to the pkg/memory/database/sqlite/db.go mentioned in the checklist).
   • The memory SQLite DB utilizes this central entry point, so the concurrent state is natively guaranteed by the connection pool settings.
   • Separate read pools were not explicitly introduced because sqliteutil.OpenDB already enforces serialized writes via SetMaxOpenConns(1) while leaving read paths unlocked. WAL seamlessly handles the read/write concurrency.

2. Cross-Platform File Lock:

   • Wrapped the entire write path (add/update/delete) inside the FileLock boundary across Linux, macOS, and Windows.

3. Drift Guard Note:

   • Since the drift guard / generation mechanism is not yet present in the current codebase, the specific test case "drift guard fires correctly under contention" cannot be fully validated in isolation at this moment. However, because the write paths are now fully guarded by the FileLock boundary, the upcoming drift guard feature can be safely integrated into this perimeter later.

All cross-platform compilation checks and go test -race for memory/database modules have passed. Ready for review!