Bug in dm_control_suite_d4pg.ipynb

The example code provided for running D4PG with tasks from DeepMind Control Suite has a bug.

The dataset preprocess function should concatenate the observations/next observations dict so that a complete observation is passed to the agent.
```

def flatten_observation(observation):
keys = sorted(six.iterkeys(observation))
observation_arrays = [tf.reshape(observation[key], [-1]) for key in keys]
return tf.concat(observation_arrays, 0)

def preprocess_fn(sample):
o_tm1, a_tm1, r_t, d_t, o_t = sample.data[:5]
o_tm1 = flatten_observation(o_tm1)
o_t = flatten_observation(o_t)
return replay_sample.ReplaySample(
info=sample.info, data=(o_tm1, a_tm1, r_t, d_t, o_t))

def main(argv):
task = dm_control_suite.ControlSuite("cartpole_swingup")

environment = task.environment
environment_spec = specs.make_environment_spec(environment)

dataset = dm_control_suite.dataset(
"./tmp",
data_path=task.data_path,
shapes=task.shapes,
uint8_features=task.uint8_features,
num_threads=1,
batch_size=256,
num_shards=100)

dataset = dataset.map(preprocess_fn).batch(256)

```