Fix the reward computation problem, step counting logic and update out-of-date dependencies (problem fix V1)

[sumwailiu]

Main changes:

• Some operations for reward-related variables are different from the Dreamer (tensorflow2 implementation) implemented by the origin author. By fixing them, I found that the Issue Does the walker run reproduce correctly? is solved, where my origin result is "episodes: 1000, total_steps: 501505, train_reward: 376.6" for the walk-run task, and the fixed result is "episodes: 1000, total_steps: 501505, train_reward: 637.5". The test reward is as follows:
  
  This well match the result from the origin paper.

• The step counting logic is not consistent, where the metrics['steps'] is computed by t * args.action_repeat during experience buffer initialization and by t during data collection. Besides, the start counting position during data collection should be 1 instead of 0. These problems make the final total_steps be 501505 = 5 (random seed episodes) * 1000 (computed by t * args.action_repeat) + 995 (the remeaning episodes) * 499 (computed by t starting from 0), which is confusing. I unify the counting rules into t * args.action_repeat to make metrics['steps'] represent the number of env interactions, and correct the start counting position for each episode during data collection.

Other changes:

• The default model_learning-rate is modified according to the HYPER PARAMETERS appendix of the origin paper Dream to Control: Learning Behaviors By latent Imagination.
• I found that the origin versions of torch, torchaudio and torchvision can not be installed. So I updated these packages, which works.
• env.py is updated to support "quadruped-run" task.