Fix/slime sglang data parallel size

[jackapbutler]

Problem

SlimeRunner._build_slime_flags() doesn't pass --sglang-data-parallel-size to slime's train.py. Slime versions after commit fa3c990 added validation at sglang_utils/arguments.py:142 that reads args.sglang_data_parallel_size — an attribute that only exists if argparse registers it via the CLI flag. Since SlimeRunner never passed it, argparse never creates the attribute, and validation crashes with AttributeError.

Additionally, the [slime] extra was missing ray as a dependency despite SlimeRunner directly invoking ray start, ray stop, and ray job submit.

Fix

• Add sglang_data_parallel_size: int = 1 field to the SlimeRunner dataclass
• Emit --sglang-data-parallel-size unconditionally in _build_slime_flags()
• Add ray to the [slime] optional dependency group in pyproject.toml

Tests

• Added two unit tests covering default and custom values for the new flag
• Full test suite passes (173 tests)
  Additionally this fix was tested for small scale RL experiments using a manual patch internally.

[lyzustc]

I think this is not needed? In our installation guide https://awslabs.github.io/agentcore-rl-toolkit/guides/slime-backend-setup/#installation, choose either option A through slime official docker, or option B through install script https://github.com/awslabs/agentcore-rl-toolkit/blob/main/src/agentcore_rl_toolkit/backends/slime/scripts/install_slime.sh, we both need to install slime's dependencies, and ray is included (https://github.com/THUDM/slime/blob/main/requirements.txt#L18)

[lyzustc]

Thanks for contribution! Codes about adding sglang data parallel size look good to me, but I think there is no need to add ray to pyproject.toml as ray is already the dependency of slime package. Also please address merge conflicts in uv.lock.

[jackapbutler]

Thanks @lyzustc, yes agreed on ray and have removed that commit and rebased on main