Convergence runs with importance sampling correction

[yfw]

Test the effect of the importance sampling correction flags (PR #174) on convergence:

• use_on_policy_kl_approximation=True
• use_importance_sampling_correction=True

[yfw]

Convergence runs here: https://wandb.ai/nvidia/grpo-dev-yifu?nw=fqj8zgnrza5