Stabilize CUDA runners and jobs

[pavlovic-ivan]

This PR introduces AWS as a platform on which runners will run. It also fixes the test-library.runs-on issue, in a way that runners don't steal jobs that does not belong to them. To put it in simple terms:

• job A deploys runner A
• job B deploys runner B
• runner A runs job B
• runner B runs job A
• if job A is finished, destruction of runner A is triggered
• once runner A is canceled, that was running job B, this is then visible as "Canceled operation" in the workflow

Issue above is now fixed with the refactored test-library.runs-on labels.

Prerequisites:

• updating/adding repo secrets
• rebase from pavlovic-ivan/ephemeral-github-runner
• rebase from pavlovic-ivan/ephemeral-github-runner-image (workflows triggered here can be canceled, images already exist in AWS)
• rebase from pavlovic-ivan/ghrunner-app-gcp and redeploy the app

@jgiannuzzi @m4rs-mt we can do a session again on this to setup everything

Working examples:

• https://github.com/pavlovic-ivan/ILGPU-fork2/actions/runs/2728293647, you can go through 14 attempts made. all of them have CUDA jobs executed successfully

Secrets to add/update:

• AWS_ACCESS_KEY_ID
• AWS_SECRET_ACCESS_KEY
• AWS_REGION
• PULUMI_BACKEND_URL