Finch HTTP/2 pools can't detect zombie connections (dead network path,
no TCP RST). On cloud infra like Fly.io, idle connections get silently
dropped, causing retries to fail repeatedly on the same dead pool.

- Add PushX.reconnect/0 to restart Finch pool with fresh connections
- Call reconnect automatically on first connection error before retry
- Enable TCP keepalive on all HTTP/2 pools for OS-level detection