Fix: Move RNAD/NFSP agent creation outside game loop

[gsmithline]

Summary

• Fixed performance bug where RNAD/NFSP agents were being re-instantiated on every game iteration
• Agents are now created once before the game loop and reused

Problem

The RNaDAgentWrapper and NFSPAgentWrapper were created inside the for gi in range(games) loop in pyspiel_runner.py. This caused:

• JAX/Haiku model re-initialization on every game
• Repeated JIT compilation overhead
• 50 games × 2 agents = 100 model loads instead of 2

On GitHub Actions (AMD EPYC CPUs), JAX compilation is slow, causing runs to take hours instead of minutes.

Solution

Move agent instantiation before the game loop so models are loaded once and reused.

Test plan

• Tested locally with docker compose - games progress smoothly
• RNAD models now load only once at startup
• Rebuild Docker image and retest on GitHub Actions

🤖 Generated with Claude Code