[RLlib] Flatten dict-typed observations before comparing them.

[0Pinky0]

When new episodes are added into replay buffers, thier first obs will be compared with the last obs from previous episode in the buffer inside concat_episode method to make sure they are same obs:

assert np.all(other.observations[0] == self.observations[-1])

Currently these obs are assumed to be pure ndarray, and dict-typed obs will raise "ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()".

This PR leverages flatten_inputs_to_1d_tensor method from rllib.utils.numpy to flatten dict-typed obs before the comparison.