Added a few features to make rl_games more comparable to the brax with jax ppo

[Denys88]

• Added Squashed tanh distribution. They use it in brax ppo. Didn't work great for me
• Added ability to return balanced results during evaluation. Unfortunately BRAX doesn't do resets with randomization.
• By default it just stores first observation and returns to it. It means if we have one env with borken start position where we always get done in a few step it overflows results. Need to use 'balance_env_rewards' in evaluation to get the true scores.
• Added 'epochs_between_resets' which is by default zero. Could be used with BRAX because they don't randomize position for envs during training in reset.
• It means that training rewards are much lower in rl-games than real one.
• Added one more example of the Simple Neural network called SimpleNet. If you use it you can get a little bit better performance. And overall it is a good example:

1. Works only for continuous.
2. [512, 256, 128] hardcoded layers single network for mu and val and independent layer for std.
3. torch.compile.
   It didn't add a lot of perf in my case but nice to have.