Mirror GPU resource slices to NVIDIA_VISIBLE_DEVICES

[v-shobhit]

Summary

• Mirror SFlow GPU resource slices into NVIDIA_VISIBLE_DEVICES alongside CUDA_VISIBLE_DEVICES
• Keep CUDA and NVIDIA container visibility aligned for srun/Pyxis tasks
• Document both env vars in GPU resource and replica docs

Test Plan

• uv run pytest tests/unit/test_app_assembly_build_task_graph.py -k "resources_gpus_count_sets_cuda_visible_devices_with_offset or resources_gpus_respects_compute_node_num_gpus"

[RogerLiu312]

Now we export both 2 envs with identical slices, see: https://github.com/NVIDIA/nv-sflow/blob/main/src/sflow/core/backend.py#L129