Retry NVIDIA containers without display capability

[peterschmidt85]

Fixes #4004.

Currently, dstack-shim asks Docker for this NVIDIA container.DeviceRequest.Capabilities set:

gpu,utility,compute,graphics,video,display,compat32

This is wrong on headless NVIDIA hosts where CUDA works but /dev/nvidia-modeset is absent. Requesting display makes NVIDIA Container Runtime fail before the user command starts:

nvidia-container-cli: mount error: stat failed: /dev/nvidia-modeset: no such file or directory

The fix keeps the current full capability set on the first start attempt. If an NVIDIA container fails specifically because /dev/nvidia-modeset is missing, the shim removes the failed container and retries without only display:

gpu,utility,compute,graphics,video,compat32

NVIDIA documents display as X11 display support. The documented default driver capabilities are utility,compute: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/docker-specialized.html

The fallback applies only to the normal NVIDIA Docker DeviceRequest path. AMD, Tenstorrent, Intel, and explicit GPUDevices handling are unchanged.

AI Assistance: This PR was prepared with AI assistance.