Reduce K8s plugin template duplication

[khluu]

Summary

• Replace three near-identical template dicts (h100_plugin_template, nebius_h200_plugin_template, a100_plugin_template) with shared constants (_COMMON_ENV, _COMMON_VOLUME_MOUNTS, _COMMON_VOLUMES) and a _build_k8s_template() builder function
• Fix bug on line 135 where DeviceType.A100.value was used instead of DeviceType.A100 (inconsistent with other comparisons and would never match since step.device is a DeviceType enum)
• Add ValueError for unsupported device types instead of silently leaving plugin as None and crashing with TypeError on line 140

Test plan

• H100: no nodeSelector, no priorityClassName, image gets pull-through-cache replacement
• A100: has nodeSelector for A100, has priorityClassName "ci"
• H200: has nodeSelector for gpu-h200-sxm, hardcoded 8 GPUs
• Unsupported device raises ValueError
• All configs include common env (VLLM_USAGE_SOURCE, HF_TOKEN secret ref)
• num_devices is correctly set in resource limits
• Two calls return independent dicts (no shared mutation)
• All 13 tests pass locally

🤖 Generated with Claude Code