Allow container runtime executable path to be specified

[elezar]

This change adds support for specifying the container runtime executable path. This can be used if, for example, there are two containerd executables and a specific one must be used.

Deploying to a k0s system with:

helm install gpu-operator -n gpu-operator --create-namespace \
  nvidia/gpu-operator $HELM_OPTIONS \
    --version=v25.3.0 \
    --set toolkit.repository=ghcr.io/nvidia \
    --set toolkit.version=ae385428-ubuntu20.04 \
    --set toolkit.env[0].name=RUNTIME_CONFIG \
    --set toolkit.env[0].value=/etc/k0s/containerd.d/nvidia.toml \
    --set toolkit.env[1].name=RUNTIME_SOCKET \
    --set toolkit.env[1].value=/run/k0s/containerd.sock \
    --set toolkit.env[2].name=RUNTIME_EXECUTABLE_PATH \
    --set toolkit.env[2].value=/var/lib/k0s/bin/containerd \
    --set toolkit.env[3].name=NVIDIA_RUNTIME_NAME \
    --set toolkit.env[3].value=nvidia

Allows the config to be extracted correctly and unblocks the deployment.

Fixes #803

[tariq1890]

Why not just executable-path or runtime-executable-path? This is consistent with how socket is named

[elezar]

Yes, I think this may be better. Let me take another pass.

[elezar]

@tariq1890 do we need to backport this?

[tariq1890]

Backporting makes sense since we are most likely releasing gpu-operator v25.3.1