Make Harbor images Prime VM GPU capable

[xeophon]

This stacks on top of #8 and makes the generated Harbor agent/verifier images capable of seeing GPUs in Prime VM sandboxes.

Changes:

• switch Harbor agent/verifier images from upstream nvidia/cuda to Prime's driver-aware CUDA VM base
• set the expected Prime GPU/NVSwitch topology for 1-GPU, no-NVSwitch PTB sandboxes
• invoke initialize-nvidia-gpu-stack from the PTB entrypoint before log streaming and monitoring starts
• skip the Fabric Manager branch for zero-NVSwitch sandboxes

Validation run on 2026-06-01:

• git diff --check -- src/harbor_adapter/template/environment/Dockerfile src/harbor_adapter/template/tests/Dockerfile src/harbor_adapter/template/environment/entrypoint.sh
• bash -n src/harbor_adapter/template/environment/entrypoint.sh
• built Prime image cmmvfch400000n3w57owbyhkq/posttrainbench-gsm8k-qwen3-1.7b-agent:pr8-c9176a7-primecuda1gpu-init-skipfm as both container and VM artifacts

GPU VM smoke is currently blocked by Prime VM sandbox startup: tiny control images on H200_141GB and RTX_PRO_6000 terminate before reaching runnable state with exit_code=1, and prime sandbox logs returns HTTP 500 for those VM sandboxes. Prime CLI was upgraded from 0.6.6 to the available 0.6.9 path and the H200 control still failed the same way.

[xeophon]

sorry, codex was a bit overeager here