Bug fixes and improvements

• fixed computation of training metrics in TRPO and PPO1
• added reset_num_timesteps keyword when calling train() to continue
  tensorboard learning curves
• reduced the size taken by tensorboard logs (added a
  full_tensorboard_log to enable full logging, which was the
  previous behavior)
• fixed image detection for tensorboard logging
• fixed ACKTR for recurrent policies
• fixed gym breaking changes
• fixed custom policy examples in the doc for DQN and DDPG
• remove gym spaces patch for equality functions
• fixed tensorflow dependency: cpu version was installed overwritting
  tensorflow-gpu when present.
• fixed a bug in traj_segment_generator (used in ppo1 and trpo)
  where new was not updated. (spotted by @junhyeokahn)