Fix critical RL reward function exploits and training hyperparameters

Align with Advanced Llama 3.2 GRPO LoRA reference notebook pattern

Fix VALID_TARGETS_FOR_ATTACK and attacker heuristic/prompt inconsistencies

Align train.py and Colab notebook with official Unsloth+OpenEnv GRPO patterns

Add multi-agent GRPO training for all 3 agents (worker, attacker, oversight)

Remove hackathon_env template, rewrite train.py for SentinelOpsArena

Initial project setup for OpenEnv Hackathon