fix: openclaw-nvidia rename, device pairing, default model

[liveaverage]

• Device pairing/bootstrap was overhauled in the sandbox UI/proxy path so the overlay is more deterministic, approvals are handled more robustly, and the dashboard is only revealed once it is actually usable.
• The nemoclaw sandbox was minimally rebranded to openclaw-nvidia externally: folder/image/start command defaults changed, while most internal compatibility naming stayed intact.
• Welcome UI URL/origin handling was hardened for Brev/proxy use, including dynamic public URL detection, host preservation, and fixes for localhost/public-host regressions.
• Provider/inference bootstrap was adjusted on the host side: key injection now updates both NVIDIA provider records, tests/docs were updated accordingly, and the default model was changed away from Kimi to Qwen.
• Logging was reduced across welcome-ui and the sandbox proxy so routine asset/bootstrap traffic is quieter, while keeping pairing and error signals visible.