test(bdd): align local GPU instance types with two-GPU topology

[sbaum1994]

Description

Follow up after #999 merges.

PR #999 intentionally changes the shared local fake GPU topology from gpuCount: 128 to gpuCount: 2 so NVCA registers the one-GPU instance type required by the LLM routing scenario. Several local k3d BDDs still deploy NCP.GPU.H100_8x, which is not compatible with the adjusted topology.

Update the local single-cluster, multi-cluster, and NVCT task-smoke BDD fixtures to select an instance type actually registered by the two-GPU topology. Keep the gpuCount: 2 adjustment from #999. Do not change the EKS BDD, which uses a real cluster topology.

Definition of Done

• Update local single-cluster Helmfile BDD deployments that request NCP.GPU.H100_8x.
• Update local multi-cluster Helmfile BDD deployments that request NCP.GPU.H100_8x.
• Update the local NVCT task-smoke default instance type.
• Preserve topology.nodePools.default.gpuCount: 2.
• Verify the selected instance type is registered before deployment.
• Run the affected local BDD scenarios and document the results.

Resources (optional)

• feat(llm-request-router): route LLM workers through backend router #999
• tools/ncp-local-cluster/apps/fake-gpu-operator/values.yaml
• tests/bdd/features/single-cluster-helmfile.feature
• tests/bdd/features/multi-cluster-helmfile.feature
• tests/bdd/scripts/run-nvct-task-smoke.sh