Added training code for PPO, A2C, DQN and random and greedy baselines

[Thomson-Lam]

Changes

• added src/models/benchmarking.py for core metric tracking used for training for all algos
• edited src/models/evaluate_agents.py as offline RL agents evaluation after training
• added a no action baseline policy inside src/models/benchmarking.py that does not suppress fires as the baseline to compare against
• edited src/models/fire_env.py to support the metrics needed according to the proposal

Main metrics:
mean_return: mean episodic return
asset_survival_rate: fraction of episodes with assets_lost == 0
containment_success_rate: fraction of episodes where the fire is extinguished before truncation
mean_burned_area_fraction: final burned-area fraction per episode, (burned + burning + asset_burned cells) / 625
std_across_seeds: standard deviation of the seed-level metric means

• made a single interface for training all RL algos, src/models/train_rl_agent.py for argparse CLI usage
• wrapped the CLI commands for training RL agents and evaluating them in Powershell and Bash inside scripts
• initialized 5 fixed seeds (11, 22, 33, 44, 55) for initializing models and evaluation order for holdout environments for benchmarking
• updated the README and docs for usage and details