better show results in tensorboard

[luoye2333]

We can use disc/loss, disc/agent_acc,.etc instead of disc_loss, disc_agent_acc. By doing so, these results will show in one group in tensorboard. This increases simplicity when checking the training results.

Also optimize the imitate rewards.

if "reward_raw" in train_info:
    reward_raw=train_info['reward_raw'].cpu().numpy().tolist()
    train_info_dict["rewards/body_pos"] =  reward_raw[0]
    train_info_dict["rewards/body_rot"] =  reward_raw[1]
    train_info_dict["rewards/lin_vel"] =  reward_raw[2]
    train_info_dict["rewards/ang_vel"] =  reward_raw[3]
    train_info_dict["rewards/power"] =  reward_raw[4]

[luoye2333]

Added two new changes in my second commit.

1. fit_smpl_shape.py
   change ths visualize process in fit_smpl_shape.py to just show the matched joints
   also added a legend for the two groups of data

2. fit_smpl_motion.py
   using Adadelta optimizer has a poor performance, better use the adam optimizer. Moreover, we can merge the two separate optimizers for dof_pos_new and root_rot_new, root_pos_offset into one.
   Also disabled the breakpoint in line 214