Implement Noisy Networks for DQN about tf2rl HOT 2 CLOSED

Implementation is done, but not has tested performance on Atari task

from tf2rl.

The result of SpaceInvader-v4 seems to be better than paper score.

• tf2rl score: ~6000 (scores on graph below are averaged over 5 run)
• paper score: 2196+92

• Reproduce the result (note the generated graph will not be identical to above one because of random seed):

# Checkout to master on 19/6/29
$ git checkout bf83a9cbcfeb14af5debb9163646c1fa8a493cc0
$ python examples/run_dqn_atari.py --enable-noisy-dqn

from tf2rl.