Migrate obs/reward normalization from env.wrappers into Agent itself

[zuoxingdong]

• Make online statistics as nn.Parameter and registered inside the module. It becomes trackable

  • Similar style with how the BatchNorm is implemented in PyTorch

• Different behavior between train/eval modes.

  • Train mode: update statistics
  • Eval mode: use current statistics without further updating

• Unit test: before replacement, benchmark with the old behavior

  • Mujoco environments: HalfCheetah, Hopper, Walker
  • Seeds: 10 seeds
  • Confirm no significant effect on performance