Support Qwen3 dense 8B and 32B true-on-policy

[maocheng23]

Summary

• add Qwen3-8B and Qwen3-32B to the dense true-on-policy model profile
• add size-specific launcher defaults for 8B and 32B, including 32B CPU optimizer offload and cuda graph batch sizes
• keep explicit parallelism/rollout overrides supported in the Qwen3 dense launcher

Testing

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest -q -o addopts= --confcutdir=tests/fast/true_on_policy tests/fast/true_on_policy/test_config.py tests/fast/true_on_policy/test_run_qwen3_4b.py
• git diff --check
• PYTHONPATH=. python -m py_compile scripts/run_qwen3_4b.py miles/true_on_policy/model_profiles.py miles/true_on_policy/schema.py

[gemini-code-assist]

Code Review

This pull request adds support for the Qwen3-8B and Qwen3-32B models. It introduces a Qwen3DenseLaunchDefaults dataclass and a lookup dictionary to manage model-specific launch configurations, such as tensor model parallel size, context parallel size, max tokens per GPU, and CPU offload options. Additionally, it updates the launch script arguments and execution logic to utilize these defaults when not explicitly provided, and adds corresponding unit tests to verify the new configurations. There are no review comments, so I have no feedback to provide.