Do not leak psycopg2 connections during testing (#628)

According to psycopg2 documentation connections are not closed on context exit.
This is in contrast with connection cursors that actually do so.

It is not an inherent problem in itself, however, calls to get_state() in a
loop, such as from wait_until_*, can overflow the connection limits and cause
the test suite to fail. This is particularly noticable when the system is slow
to respond, for example if it is run under valgrind.

This might address other objerved flakyness in the tests due to connection
exhaustion with messages as:
"psycopg2.OperationalError: FATAL:  sorry, too many clients already"

Co-authored-by: Georgios Kokolatos <gkokolatos@pm.com>

Author: gkokolatos

tests/pgautofailover_utils.py

@@ -359,14 +359,21 @@ def run_sql_query(self, query, *args):
         Runs the given sql query with the given arguments in this postgres node
         and returns the results. Returns None if there are no results to fetch.
         """
-        with psycopg2.connect(self.connection_string()) as conn:
-            cur = conn.cursor()
-            cur.execute(query, args)
-            try:
-                result = cur.fetchall()
-                return result
-            except psycopg2.ProgrammingError:
-                return None
+        result = None
+        conn = psycopg2.connect(self.connection_string())
+
+        with conn:
+            with conn.cursor() as cur:
+                cur.execute(query, args)
+                try:
+                    result = cur.fetchall()
+                except psycopg2.ProgrammingError:
+                    pass
+        # leaving contexts closes the cursor, however
+        # leaving contexts doesn't close the connection
+        conn.close()
+
+        return result
 
     def pg_config_get(self, settings):
         """