CI: de-DINDify

The current DIND-based CI practically limits the host distribution to Debian userspace + irregular system configuration (systemd, udevd, resolved, etc.) + irregular kernel (Ubuntu, or Oracle on Lima workflows).
https://github.com/moby/moby/blob/8210b0babaed17fbe706a4c5e8c7d0b9c6aba2a1/Dockerfile#L4
https://github.com/moby/moby/blob/8210b0babaed17fbe706a4c5e8c7d0b9c6aba2a1/.github/workflows/test.yml#L35
https://github.com/moby/moby/blob/8210b0babaed17fbe706a4c5e8c7d0b9c6aba2a1/.github/workflows/vm.yml#L46-L47

To cover "real" host OSes with their standard configuration, we have to consider de-DINDifying the CI.
I suggest beginning with de-DINDIfying ./integration/container ./integration/build ./integration/system executed in vm.yml.

openSUSE's test script might be useful for reference https://github.com/os-autoinst/os-autoinst-distri-opensuse/blob/master/tests/containers/docker_engine.pm

• .vm.yml #53672
• .test.yml
  • integration
  • integration-cli
  • integration-flaky
  • docker-py

Start with .github/workflows/vm.yml and the ./integration/container, ./integration/build, and ./integration/system targets named in the issue. Compare the current DIND-based execution with the openSUSE test script for reference. Done means these integration areas run in vm.yml without relying on Docker-in-Docker while covering hosts with their standard configuration.