Recover the connection when a second Ctrl-C interrupts the query cancel

[DiegoDAF]

Description

Pressing Ctrl-C twice on a long query can leave the connection unusable: every later statement, rollback included, fails with

sending query failed: another command is already in progress

and the query keeps running on the server. We hit it over an SSH tunnel to a remote server, but a tunnel is not required, just enough round-trip time.

psycopg's Connection.wait() answers the first Ctrl-C by sending a cancel and then reading the server's reply. A second Ctrl-C during that cleanup interrupts it: the cancel may never go out and the reply is never read. execute_command() then prints "cancelled query" without looking at the connection.

After a cancelled query, pgcli now checks whether the connection is still busy. If it is, it cancels again and reads the pending reply, using psycopg.pq only and waiting at most 5 seconds. If the server does not answer, or yet another Ctrl-C arrives, it cancels (at most 2 seconds, ignoring Ctrl-C meanwhile so the query does not outlive the session) and closes the connection, so the next statement goes through the existing reconnect path. A connection that is not busy is left alone, so session state survives an ordinary Ctrl-C.

Reproduced end to end by driving pgcli with pexpect through a TCP proxy that adds 150 ms each way: two Ctrl-C 0.15 s or 0.4 s apart left the connection busy and the pg_sleep running on the server. With this change the next statement works and the sleep is cancelled, also with three, four or six quick Ctrl-C.

5 tests against a live server, all failing on main; two of them also fail if the last-resort path only closes the connection, or lets a Ctrl-C skip its cancel.

The cleanup gap itself is in psycopg, reported as psycopg/psycopg#1444. A fix there would leave the connection closed instead of busy; this change is still needed to cancel the query and keep the session when the connection can be recovered.

Checklist

• I have added this contribution to the changelog.rst.
• I have added my name to the AUTHORS file (or it is already there).
• I installed pre-commit hooks (pip install pre-commit && pre-commit install).
• Please squash merge this pull request (uncheck if you would like us to merge as multiple commits)

On the third box: .pre-commit-config.yaml pins an older ruff-pre-commit than the one tox -e style installs, and the two disagree on formatting, so I ran ruff check and ruff format matching the CI instead, both clean.