fix libpq nonblocking connect readiness

[russellromney]

Summary

Fix two TCP readiness issues that break libpq/psycopg's nonblocking connect path under Quark:

• Clear the select(2) exception fdset entry itself when no exception event occurred.
• When uring TCP getsockopt(SO_ERROR) only has cached EINPROGRESS, ask the host socket for the current SO_ERROR; if the host reports success, clear the cached state and promote the socket through PostConnect().

Root cause

psycopg/libpq opens the TCP socket nonblocking, receives EINPROGRESS from connect(2), and then immediately checks getsockopt(SO_ERROR) before sending the PostgreSQL startup packet. Quark was still returning cached EINPROGRESS until a later readiness path advanced the uring socket state.

A separate select fdset typo caused sockets to appear in the exception fdset even after SO_ERROR was clear.

Validation

Validated through Tinyhost's real droplet e2e against postgres:16-alpine:

SOCKET connect_ex=115 immediate_so_error=0 select_read=False select_write=True select_except=False so_error=0
RESULT quark/socket PASS
RESULT quark/pg8000 PASS
RESULT quark/psycopg PASS
root=clear; quark libpq nonblocking connect works against real Postgres
62_QUARK_PSYCOPG_OK