Runner requires manual call to close() so calling-code can work with "dirty" runner

[lefnire]

A conversation I'd had earlier with @AlexKuhnle - in order to do cross-validation, or alternating train/test, we came to doing this:

# Train
runner = Runner(agent=agent, environment=env)
runner.run(timesteps=...)

# Test
next_state, terminal = env.reset(), False
while not terminal:
    next_state, terminal, reward = env.execute(agent.act(next_state, deterministic=True))

With the current TensorForce code, the Test segment will fail because agent & model will have been closed automatically (in the end of runner.run()). For consideration: this PR which requires manually closing runner, or possibly a method argument such as auto_close=True.