docs: update GPU operator guide with Civo-tuned flags and H100 NVLink…

[jokestax]

Summary

Updates the GPU Clusters doc (kubernetes/advanced/gpu-config.md) so customers get a reliable install on first try on Civo's GPU image, and have clear guidance for the single-GPU H100 case that currently fails without a workaround.

What changed

• Install command tuned for Civo's image. Replaced the bare helm install --generate-name snippet with a helm upgrade --install that passes toolkit.enabled=false (the NVIDIA container toolkit is already baked into Civo's GPU image), plus driver.enabled=true, devicePlugin.enabled=true, gfd.enabled=true, operator.defaultRuntime=containerd, and validator.cuda.runtimeClassName=nvidia. Added a table explaining what each flag does.
• New section: Single-GPU H100 nodes — NVLink workaround. On a node with only one H100, the driver fails to load because NVLink has no peer. Documented the nvidia-kernel-config ConfigMap with NVreg_NvLinkDisable=1 and the matching driver.kernelModuleConfig.name helm flag. Includes a warning not to apply this on multi-H100 nodes.
• B200 added to the supported GPU list.
• GPU Operator version pinned — called out chart/app v25.10.1 as the version Civo has validated end-to-end, with a --version 25.10.1 tip.
• Troubleshooting expanded — H100 CrashLoopBackOff now points at the NVLink section; added guidance for pods stuck Pending on nvidia.com/gpu.

Why

The previous install example produced a working-but-suboptimal setup on Civo clusters (the Operator would layer its own container toolkit on top of the one Civo already provides), and single-GPU H100 nodes failed with no customer-facing explanation. These edits bring the doc in line with what our internal testing has validated.

[hlts2]

LGTM