#6570 Restore Stable PostgreSQL I/O and reduce scheduler work

Addresses the correctness and scheduler findings in #6568, on top of #6532.

• Restore Stable's original PostgreSQL poll-time I/O and handler API; gate deferred readiness and TLS handshake caching with PROXYSQL31.
• Preserve queued output on retryable plaintext send failures on both tiers.
• Process backend responses before new frontend work, skip unnecessary late scans, and publish only the requested MySQL hostgroup's cached connections.
• Batch terminal continuation cleanup and dispatch capacity released by deletion. Add socket, ordering, cache-scope, paused-session and cleanup regressions.

Validation: Stable and v3.1 release/debug builds passed. The v3.1 release and debug suites each passed 261 assertions; focused Stable checks and real native/libpq waiter-ordering checks passed. Throughput improvement has not been established.

---

Summary by cubic

Restores Stable's original PostgreSQL poll-time I/O and handler API, and reduces scheduler work by processing backend responses first and skipping unnecessary scans.

• Deferred readiness and TLS handshake caching stay gated behind PROXYSQL31; Stable keeps its original socket I/O in the poll scan.
• Retryable plaintext send failures now preserve queued output on both MySQL and PostgreSQL tiers instead of closing healthy slow clients.
• Backend responses are dispatched before new frontend work; late scans are skipped when no backend work is pending.
• Normal checkout recovery publishes only the requested MySQL hostgroup's cached connections, leaving compatible caches for other hostgroups intact.
• Terminal continuation cleanup is batched into a single traversal, with released capacity dispatched to waiters in the same pass.
• Adds real-socket and scheduler-cost regression tests covering poll boundaries, paused disconnects, retryable sends, targeted cache publication, and cleanup ordering.

Summary by CodeRabbit

• Bug Fixes
  • Cached MySQL connections are now returned only for the requested hostgroup, helping prevent connections from unrelated hostgroups being used.
  • PostgreSQL connections better handle temporary send interruptions: queued data is retained for retry, while fatal send errors close the stream.
  • Worker scheduling now performs cleanup and capacity handling more consistently, including for paused or terminated sessions.
• Compatibility
  • PostgreSQL polling and session handling retain build-specific behavior across Stable and ProxySQL 3.1 configurations.