[skyrl-train] Upgrade vllm to 0.13.0 (and torch to 2.9.0) for both vllm and mcore extras

[erictang000]

Upgrading vllm to latest (minor changes).

Needed for #885 and should subsume #882

Upgrades transformer-engine from 2.9.0 -> 2.10.0 for the megatron backend due to incompatibility with triton 3.5 (required by torch 2.9.0)

Keeps the vllm_engine.py path backwards compatible for versions of vllm < 0.11.2 (important for flash-rl integration)

[gemini-code-assist]

Code Review

This pull request upgrades vllm to version 0.13.0 and torch to 2.9.0, which is a necessary update. The changes include updating dependencies in pyproject.toml and adapting the Python code in vllm_engine.py to accommodate breaking changes in the new vllm version. The changes are generally good, but I have one suggestion to improve the readability and robustness of the API compatibility logic.

[gemini-code-assist]

The logic to handle backward compatibility for different vllm API versions is functionally correct but could be simplified for better readability and maintainability. By grouping the version-specific logic, the code becomes easier to follow.

Additionally, the else block for OpenAIServingModels instantiation was changed to use keyword arguments. While this can improve explicitness, it's safer to retain the original positional arguments for the older API path to guarantee backward compatibility, as the function signature for that version might not support keyword arguments for these parameters.

This suggested change refactors the logic to be more grouped and uses a safer approach for the legacy path.

Suggested change

	# vllm >= 0.11.2 removed model_config from OpenAI serving APIs
	is_new_api = version.parse(vllm.__version__) >= version.parse("0.11.2")
	legacy_kwargs = {} if is_new_api else {"model_config": model_config}
	if is_new_api:
	models = OpenAIServingModels(engine, base_model_paths)
	else:
	models = OpenAIServingModels(engine, model_config=model_config, base_model_paths=base_model_paths)
	# vllm >= 0.11.2 removed model_config from OpenAI serving APIs
	is_new_api = version.parse(vllm.__version__) >= version.parse("0.11.2")
	legacy_kwargs = {}
	if is_new_api:
	models = OpenAIServingModels(engine, base_model_paths)
	else:
	models = OpenAIServingModels(engine, model_config, base_model_paths)
	legacy_kwargs["model_config"] = model_config