Add GSM8KEnv and GSM8KAgent for Tunix agentic RL

[andytwigg]

Summary

This PR introduces a modular GSM8KEnv environment, GSM8KAgent agent, and an end-to-end training entrypoint script train_gsm8k_qwen3.py for Qwen3-1.7B for the Tunix Agentic RL framework, matching the architecture and conventions established in examples/frozenlake/.

Changes Included

1. examples/math_gsm8k/env.py:

   • GSM8KEnv: Subclasses BaseTaskEnv to manage mathematical reasoning tasks.
   • Extracts ground-truth numerical answers from GSM8K solutions via extract_hash_answer (... #### <num>).
   • Parses model reasoning chains and LaTeX \boxed{...} answers using stack-based bracket matching with robust fallbacks.
   • Computes composite format and accuracy rewards (+0.1 format bonus, +1.0 exact accuracy, +0.5 partial credit).
   • Provides serialization and dynamic instantiation via from_dict().

2. examples/math_gsm8k/agent.py:

   • GSM8KAgent: Subclasses ConversationAgentBase to maintain dialogue history and step-by-step reasoning trajectories for GRPO training.

3. examples/math_gsm8k/data.py:

   • GSM8K dataset loaders supporting Hugging Face (openai/gsm8k), TFDS, and in-memory smoke-testing.

4. examples/math_gsm8k/env_test.py:

   • 13 comprehensive unit tests validating answer extraction, tag checking, reward calculation, and agent-environment interaction.

5. examples/math_gsm8k/train_gsm8k_qwen3.py:

   • Agentic GRPO training entrypoint script for GSM8K on TPU using Qwen3-1.7B.
   • Integrates GSM8KEnv and GSM8KAgent with GRPOLearner and RLEngine.
   • Configures ModelConfig.qwen3_1p7b() with bf16 reference and fp32 actor (supporting full fine-tuning and LoRA).
   • Supports vLLM and vanilla rollout engines on shared or distributed mesh.
   • Computes and logs step-by-step math reasoning metrics (solve ratio, format ratio, reward).

TAG=agy
CONV=2907f211-714a-4e9c-b009-de21347c291a