Improve pool connection handling

[nils-borrmann-tacto]

I think that the current implementation of the connection pool is not ideal, which leads to some subtle problems.

The connection pool always "refreshes" every connection it takes from the pool, and thereby simply reconnects a connection that had been closed. This happens even though there might be another still alive connection in the pool.

Imagine a situation where the pool has two free_connections. The first one is closed, the second on still alive. I would expect the connection pool to test all connections until it finds one that works and acquires that connection. This removes closed connections from the pool and avoids additional latency from unnecessarily re-establishing a new connection.

I therefore changed the behaviour of the connection acquisition in the pool. The pool no longer "refreshes" connections, but only checks if the connection is still live (by sending a ping). It does so until it finds a live connection.

(There was actually another bug in the refreshing: Re-connecting the connection never actually worked, because the check for if self._opened in connect() always returned True, thus exiting the connect method. This only worked because the ExecutionContext inside execute() finally re-established the connection.)

I also added some debug logs, which helps understanding this behaviour.

[nils-borrmann-tacto]

I think there was a race condition lingering here. Directly after our connection was created, it could be acquired by another coroutine. I fixed this by never putting the connection into the free pool.

[nils-borrmann-tacto]

This avoid an exception when closing an already closed connection.