Fix Gym Env and Implement RL Training

[tztsai]

• Fixes gym.py in psl to use BuildingEnv to wrap any BuildingEnvelope system (the original version wraps any instance of ODE_NonAutonomous, but in its implementation, it seems to assume that the system is a BuildingEnvelope).
• Implements two DRL algorithms ppo.py and sac.py (largely adopted from https://github.com/vwxyzjn/cleanrl) in the rl folder. Both DRL algorithms can successfully run in the BuildingEnv environment.
• Implements gym_dpc.py and gym_nssm.py where a DPCTrainer or NSSMTrainer can directly accept a gym environment as input and use a neuromancer.Trainer to train a neural network in this environment.
• Drafts hybrid_control.py as an attempt to implement the technical proposal in README.md, illustrated by diagram.svg. Currently the program can successfully train a DPC policy in a BuildingEnv, and insert its policy model as the actor network of an actor-critic PPO agent, and then continue the training in an RL workflow. This hybrid approach may have the benefit of improving the DPC policy by learning from long-term cumulative reward, as well as accelerating the DRL training by providing a pre-trained DPC policy model.