[Feature Request] Add minimal 1-GPU script for debugging and learning

Motivation

All existing run scripts under scripts/ require 4-8 GPUs (e.g., run-qwen2.5-0.5B-reproducibility.sh uses 8 GPUs, test scripts use 4 GPUs). This makes it difficult for new contributors to:

• Understand the slime training flow on a single development GPU
• Debug issues without reserving a multi-GPU node
• Quickly iterate when developing new features (e.g., adding model support)

Proposal

Add a minimal 1-GPU run script scripts/run-qwen3-0.6B-minimal.sh that:

• Uses Qwen3-0.6B (smallest supported Qwen3 model)
• Runs on a single GPU with --tensor-model-parallel-size 1, --pipeline-model-parallel-size 1
• Uses small batch sizes (global batch size 16, rollout batch size 4) and short response lengths (512 tokens)
• Covers the full GRPO training loop (rollout + training + eval) on GSM8K

Existing alternatives

None of these provide a simple 1-GPU script for the standard training flow.

Additional Context

I'm happy to submit a PR for this if the team thinks it would be useful. The script is a single file with no changes to existing code.

Compare scripts/run-qwen2.5-0.5B-reproducibility.sh with examples/true_on_policy/run_simple.py and tests/test_qwen2.5_0.5B_debug_rollout_then_train.py to identify the standard training flow and required invocation. Add scripts/run-qwen3-0.6B-minimal.sh with the listed one-GPU, batch-size, response-length, and GSM8K settings. Done means rollout, training, and evaluation complete on one GPU.