DDPG bug: layer norm not really applied when initializing the critic (Q) network

[xuanlinli17]

In the DDPG implementation, in models.py, note that the **network_kwargs in
self.network_builder = get_network_builder(network)(**network_kwargs)
does not contain layer_norm=True/False. Thus, when the critic uses this network builder to build mlp, layer norm is not applied. This causes the model to fail on many environments such as HalfCheetah.

Variable names of the critic in the original code:
critic/mlp_fc0/w:0
critic/mlp_fc0/b:0
critic/mlp_fc1/w:0
critic/mlp_fc1/b:0
critic/output/kernel:0
critic/output/bias:0

Variable names of the critic should be:
critic/mlp_fc0/w:0
critic/mlp_fc0/b:0
critic/LayerNorm/beta:0
critic/LayerNorm/gamma:0
critic/mlp_fc1/w:0
critic/mlp_fc1/b:0
critic/LayerNorm_1/beta:0
critic/LayerNorm_1/gamma:0
critic/output/kernel:0
critic/output/bias:0

However, even after fixing this, DDPG still runs poorly on HalfCheetah after 2M time steps (reward is less than 1000). It should reach a reward of ~3000+ according to many papers. It is possible that there are other bugs.

[DanielTakeshi]

@lilililiiiii Can you try after the change proposed in #938?