Support AWS G7e (RTXPRO6000) instances

[jvstme]

$ dstack offer -b aws --gpu RTXPRO6000 --group-by gpu,backend,region                     

 #  GPU                   SPOT             $/GPU          BACKEND  REGION      
 1  RTXPRO6000:96GB:1..8  spot, on-demand  0.5245..4.143  aws      us-east-2   
 2  RTXPRO6000:96GB:1..8  spot, on-demand  0.6513..4.143  aws      us-west-2   
 3  RTXPRO6000:96GB:1..8  spot, on-demand  1.0129..4.143  aws      us-east-1

Tested:

• Spot and on-demand
• Single-GPU and multi-GPU
• nvidia-smi
• PyTorch

[peterschmidt85]

Lets double-check that EFA works (or at least theoretically from the code) for

g7e.8xlarge
g7e.12xlarge
g7e.24xlarge
g7e.48xlarge