Fix n_updates counting with early stopping in MaskablePPO and RecurrentPPO

[alektebel]

Description

This PR fixes incorrect n_updates counting in both MaskablePPO and RecurrentPPO when early stopping is triggered due to the target_kl threshold.

The Problem:

• Previously, n_updates was always incremented by the full n_epochs value (default: 10) regardless of actual epochs completed
• When target_kl triggered early stopping (e.g., at epoch 5), n_updates would still increment by 10 instead of 5
• This caused inaccurate learning rate scheduling, progress tracking, and inconsistent behavior with regular PPO

The Solution:

• Move n_updates increment inside the epoch loop to count only completed epochs
• Each epoch now increments n_updates by 1, matching the behavior of base PPO
• Early stopping now correctly reflects actual training progress

Files Changed:

• sb3_contrib/ppo_mask/ppo_mask.py
• sb3_contrib/ppo_recurrent/ppo_recurrent.py

Testing

The fix has been verified with:

• Reproduction cases showing the bug in both algorithms
• Logs demonstrating correct counting after the fix
• Early stopping at various steps (0-7) to validate accurate incrementing

Impact

• Accurate learning rate scheduling
• Proper training progress tracking
• Consistent behavior with Stable-Baselines3 PPO
• No breaking changes - only affects counting logic