Fix flaky newly migrated MI300 workflow

[yamiyysu]

Newly migrated mi_300 workflow fails 20% of time](https://github.com/iree-org/iree/actions/runs/13144343574/job/36679240037) due to apt-installing not able to reach remote host. Propose to move the dependency in the runner scale set so that

• remove apt-install in every run to increase robustness
• remove docker-in-docker, which requires the runner scale set to run in privileged mode and can access all GPUs

This depends on first creating a new runner scale set, then migrate this workflow

[ScottTodd]

Do we know why the network is unstable? Can we fix that too? I doubt this will be the last time someone tries to install a package on these machines.

[yamiyysu]

We don't know: I haven't dug deeper to see if it client-side/Conductor, or server side, or in-between. We are in general having Conductor networking issue though.