[Feature]: Support fractional G4 GPUs on GCP

[peterschmidt85]

Problem

GCP provides fractional RTX PRO 6000 GPUs through these G4 machine types:

• g4-standard-6: 1/8 GPU, 12GB GPU memory
• g4-standard-12: 1/4 GPU, 24GB GPU memory
• g4-standard-24: 1/2 GPU, 48GB GPU memory

They are not currently supported by dstack and should be excluded as part of #4039. Unlike full-GPU G4 instances, these shapes expose an NVIDIA vGPU and require GCP's specific vGPU/GRID guest driver and VM identity authentication.

Proposed support

Expose each fractional shape as one CUDA-visible RTX PRO 6000 vGPU with its effective memory (12GB, 24GB, or 48GB), so normal dstack GPU-memory matching remains correct.

Provisioning must:

1. Select the fractional-G4 path by GCP machine type, not by globally classifying RTXPRO6000 as requiring proprietary modules.
2. Install/use Google's supported G4 vGPU/GRID driver and preserve the service account/scopes required for VM identity authentication.
3. Verify host GPU discovery before accepting the instance.
4. Keep the existing open-driver path for full-GPU g4-standard-48 and larger shapes.

Acceptance criteria

• Offers report one GPU with the effective memory for each fractional shape.
• A memory requirement cannot match a fractional shape with less effective memory.
• A minimal CUDA container sees /dev/nvidia0, a working nvidia-smi, and the expected memory.
• Full-GPU G4 offers and workloads remain unchanged.

References:

• https://cloud.google.com/compute/docs/accelerator-optimized-machines#g4-series
• https://cloud.google.com/compute/docs/gpus/install-drivers-gpu#fractional-vms

[peterschmidt85]

I personally think there is very low value in supporting fractions. Low value but need to maintain it. Azure example is quite illustrative. I'd rather drop Azure's fractional A10 too.