Make BenchmarkConfig.compute support instance configuration in YAML

[R-Palazzo]

Problem Description

The BenchmarkConfig supports defining and validating the configuration used to launch benchmark runs on remote compute services.

Currently, the compute section only requires a service key, with values such as gcp or aws. Based on that service, the instance configuration is resolved from Python-side defaults that are currently hardcoded in config_utils.py

We want to update the compute section so that instance-level configuration is defined directly in the YAML file.
As part of this change, we should define provider-agnostic terminology for the YAML schema and continue supporting provider-specific aliases internally. This will make it clearer which values users are expected to provide while still allowing the implementation to map those fields to each compute service.

Expected behavior

• Move the current configuration defined in config_utils.py to the benchmark_base.yaml:
• Define the compute instance settings explicitly in YAML, for example:

compute:
  service: gcp
  instance_type: n1-highmem-16
  boot_image: projects/deeplearning-platform-release/global/images/family/common-cu128-ubuntu-2204-nvidia-570
  root_disk_gb: 300
  gpu_type: nvidia-tesla-t4
  gpu_count: 1
  swap_gb: 64
  name_prefix: sdgym-run

• Validate the new compute schema as part of BenchmarkConfig
• Update BenchmarkLauncher and benchmark methods to work with the new structure.

Additional context

The config should use canonical keys such as:

• instance_type
• boot_image
• root_disk_gb
• gpu_type
• gpu_count

These keys can then be translated internally to provider-specific names. For example:

• instance_type → machine_type on GCP, instance_type on AWS
• boot_image → source_image on GCP, ami on AWS
• root_disk_gb → disk_size_gb on GCP, volume_size_gb on AWS