GCP DWS/MIG launches create STANDARD VMs instead of FLEX_START VMs

SkyPilot’s GCP DWS VM path uses gcp.managed_instance_group and creates a Managed Instance Group resize request with requestedRunDuration, but the regional instance template does not explicitly set the VM provisioning model to FLEX_START.

As a result, the created GCE instances can show:

scheduling:
  provisioningModel: STANDARD

instead of:

scheduling:
  provisioningModel: FLEX_START

This is surprising because the SkyPilot docs describe gcp.managed_instance_group as the DWS path for GCP GPU VMs. If the VM remains STANDARD, it may not receive Flex-start/DWS semantics or billing.

Current behavior:

sky/provision/gcp/mig_utils.py:create_region_instance_template() removes the SkyPilot-only managed-instance-group config and passes the remaining node config into the regional instance template. The existing scheduling config only rejects SPOT; it does not inject FLEX_START.

Expected behavior:

When gcp.managed_instance_group is set, SkyPilot should create the instance template with Flex-start scheduling, for example:

scheduling:
  provisioningModel: FLEX_START
  instanceTerminationAction: DELETE
  onHostMaintenance: TERMINATE
  maxRunDuration:
    seconds: <run_duration>

It should also continue using NO_RESERVATION, as required for Flex-start.