feat: default model to Qwen3.6-35B-A3B + GPU Operator v26.3 + H100

[Hadar301]

Summary

• Upgrade NVIDIA GPU Operator default from v25.3 → v26.3 (v26.3.3 validated on a fresh ROSA HCP cluster).
• Switch default GPU target from L40S (g6e.2xlarge) to H100 (p5.4xlarge) in Terraform and Helm defaults; strip redundant ROSA/ARO value overrides that now match the base, and document g6e.12xlarge (4× L40S, tensor-parallel-size=4) as an alternative.
• Use the custom vLLM runtime (v0.19.0) as the default for all clouds, replacing the RHOAI-bundled runtime. Promotes the validated ARO stack to the base.

Test plan

• helm template charts/pca-ai-serving -f charts/pca-ai-serving/values-rosa.yaml — confirmed model ID, served alias, max-model-len=262144, vllm-openai:v0.19.0, H100 resource sizing, extraEnv, tensor-parallel-size=1 all render correctly
• helm template charts/pca-ai-serving -f charts/pca-ai-serving/values-aro.yaml — confirmed ARO overlay still renders correctly with its gpuProduct override (NVIDIA-H100-NVL)
• GPU Operator v26.3.3 installed successfully on a fresh ROSA HCP 4.21.7 cluster (CSV in Succeeded phase)
• ArgoCD App-of-Apps from feat/defaults branch resolved and synced all child apps on the test cluster
• pca-operators Synced/Healthy — GPU Operator, RHOAI, DevSpaces, cert-manager all installed
• pca-devspaces / pca-platform-config OutOfSync is controller-drift only (normal), not caused by these changes
• Live smoke test against running model — blocked by p5.4xlarge provisioning issue on ROSA HCP sandbox (ROSA's management cluster does not issue EC2 RunInstances for p5 instances; ARO cluster with H100 already validates the same model/runtime stack in production)