[bug] Resolve session.close() races

[blast-hardcheese]

Why

When switching over to _terminating_task being authoritative, both _closing_waiter's state got out of sync and we did not actually set _terminating_task, which ensured N callbacks to close_session_callback instead of at most once.

What changed

• Set _terminating_task
• Rely on that instead of _closing_waiter

Test plan

test_reconnect was just a duplicate of test_connect, converting that over to handle the racing close case.