fix(vllm-performance): propagate package dependencies into Ray worker runtime environment

[mgazz]

This pull request adds propagation of the ray package dependency from the job environment into the Ray worker's runtime environment.

Previously, only ado-vllm-performance and ado-core were extracted from the job environment; now ray itself is also extracted and explicitly included in the worker dependencies, ensuring version consistency between the job environment and the Ray workers.

Fixes #1257

[AlessandroPomponio]

You're now starting to get a multitude of packages to be passed down to the runtime env of the task. I believe you might also need to pass down the vllm version (see #1214), so I'm wondering whether it's best to drop this per-package approach and just merge the dictionaries and optionally guard against passing down certain packages (basically having an opt-out approach instead of opt-in)

[mgazz]

Following @AlessandroPomponio suggestion I went for a compromise. I kept the same control for target packages (ray, vllm, guidellm) and generalised the way we set dependencies in the runtime environment.

[mgazz]

Setting the following in the remote.yaml

  fromPyPI:
    - ray==2.55.1
    - vllm[bench]==0.24.0

Generates a test environment with the following dependencies

(VLLMPerformanceTest pid=13618, ip=10.129.26.19) 2026-07-23 08:21:03,246 DEBUG AsyncIO Thread: default ado_actuators.vllm_performance.actuator: submit : Experiment (test-agentic-tool-calling) - Ray task environment: {'uv': ['${RAY_RUNTIME_ENV_CREATE_WORKING_DIR}/ado_core-2.0.1.dev44+g4d2dce39.d20260723151103.dirty-py3-none-any.whl', '${RAY_RUNTIME_ENV_CREATE_WORKING_DIR}/ado_vllm_performance-1.9.7.dev42+g4d2dce39.d20260723151103.dirty-py3-none-any.whl[vllm]', 'ray==2.55.1', 'vllm[bench]==0.24.0']}

Changes have been addressed