fix: defer EventLoopThread teardown off DbLoop to avoid self-join

[wangshuo-wsws]

Summary

Fixes std::system_error / “Resource deadlock avoided” when the last DbClient / SQLite connection / PgListener reference is released from inside its own EventLoopThread callback (#2552).

~EventLoopThread calls thread_.join(). If that destructor runs on the same worker thread (via a completion lambda holding the last shared_ptr), join() throws and the process aborts.

• Heap-allocate EventLoopThread / EventLoopThreadPool so ownership can be moved off the loop thread
• When destroying from isInLoopThread(), hand the handle to a short-lived helper thread that performs the real teardown and join()

Touches DbClientImpl, Sqlite3Connection, and PgListener (same hazard on each).

Test plan

• Run the SQLite :memory: reproducer from EventLoopThread self-join in DbClientImpl teardown when last ref released from own DbLoop thread (throws Resource deadlock avoided) #2552 in a loop; process should no longer abort
• Normal DbClient construct / execSql / destroy from a non-loop thread still works
• Postgres path: create client, run a transaction, release from a result callback without aborting

Closes #2552

[an-tao]

@harleywilsoneng Thanks so much for your patch. I've made a PR an-tao/trantor#408 for that issue, please check it out.