Support Runpod Instant Clusters

[r4victor]

#2802

The PR adds support for Runpod Instant clusters. dstack can now run multi-node tasks (2-8 nodes) on Runpod.

Implementation details:

• Added a new Compute mixin ComputeWithGroupProvisioningSupport for backends that can provision/delete multiple instances at once.
• Implemented ComputeWithGroupProvisioningSupport for RunPod.
• Added ComputeGroupModel that's used to terminate the cluster when all instances become terminating.
• process_submitted_jobs now can provision multiple jobs for multi-node tasks at once if the backend supports it (when provisioning master job). Introduced JobModel.waiting_master_job to avoid race conditions with one-by-one jobs processing.

Notes:

• Tested only 2x H100:8 and 2x A100:8 clusters since others were not available at the time.
• NCCL tests work but the bandwidth is suboptimal. This is a Runpod internal issue and they are on it.

TODO:

• templated_id is currently required in the Runpod API when creating cluster. Waiting for Runpod to make it optional. Currently pytorch template is hardcoded.