Try to fix some of the DROP DATABASE flakes (#7907)

I think one of the sources of trouble might be that _early_introspection
is hanging on to connections and can't be forced to drop them.
Try to work around it by forcing them closed at the end of introspecting.

Also, when in dev/test mode, dump out all the info about the other
connections.  Hopefully this will let us figure out where they are
coming from.

edb/server/tenant.py

@@ -993,7 +993,7 @@ async def _pg_ensure_database_not_connected(self, dbname: str) -> None:
             conns = await pgcon.sql_fetch_col(
                 b"""
                 SELECT
-                    pid
+                    row_to_json(pg_stat_activity)
                 FROM
                     pg_stat_activity
                 WHERE
@@ -1003,8 +1003,14 @@ async def _pg_ensure_database_not_connected(self, dbname: str) -> None:
             )
 
         if conns:
+            debug_info = ""
+            if self.server.in_dev_mode() or self.server.in_test_mode():
+                jconns = [json.loads(conn) for conn in conns]
+                debug_info = ": " + json.dumps(jconns)
+
             raise errors.ExecutionError(
-                f"database branch {dbname!r} is being accessed by other users"
+                f"database branch {dbname!r} is being accessed by "
+                f"other users{debug_info}"
             )
 
     @contextlib.asynccontextmanager
@@ -1222,6 +1228,18 @@ async def _early_introspect_db(self, dbname: str) -> None:
                         early=True,
                     )
 
+        # Early introspection runs *before* we start accepting tasks.
+        # This means that if we are one of multiple frontends, and we
+        # get a ensure-database-not-used message, we aren't able to
+        # handle it. This can result in us hanging onto a connection
+        # that another frontend wants to get rid of.
+        #
+        # We still want to use the pool, though, since it limits our
+        # connections in the way we want.
+        #
+        # Hack around this by pruning the connection ourself.
+        await self._pg_pool.prune_inactive_connections(dbname)
+
     async def _introspect_dbs(self) -> None:
         async with self.use_sys_pgcon() as syscon:
             dbnames = await self._server.get_dbnames(syscon)