Update Cuda runtime from 12.8 to 13.0 in dstack base Docker images

[r4victor]

Closes #3998

• Update Cuda runtime from 12.8 to 13.0 in dstack base Docker images by building from nvidia/cuda:13.0.3.
• Update NCCL_VERSION=2.28.3-1
• Pins a new dstack base Docker image version 0.15.

Cuda 13.0 requires nvidia host driver >= 580. Many dstack backends where dstack does not control the host drivders (e.g. Lambda, Crusoe, VastAI, Runpod) may provision hosts with older driver versions. But Cuda 13.0 still works since nvidia/cuda:13.0.3 includes cuda-compat-13 providing compatibility for all drivers >= 535. See https://docs.nvidia.com/deploy/cuda-compatibility/forward-compatibility.html. Previous dstack Docker images didn't include cuda-compat-13 so Cuda 13 workloads didn't work out-of-the-box on pre 580 hosts. Now they work.

Backward incompatible changes

• V100 provisioning is still supported on some backends (e.g. GCP) but requires a custom Docker image or post-installing a compatible (<13) Cuda runtime inside run commands as Cuda 13.0 dropped support for Volta and older generations.
• Drop P100 provisioning from GCP and OCI.