CI: Avoid 32 bits failures and improve behavior when failures happen

[bcardiff]

It seems the CI is able to run the std_specs for 32 bits fine enough when using a released compiler. There is some heavy memory consuming specs. But building the compiler is the thing that often fails. For example when running the run macro to embed tool init .ecr templates. (See #8279 for a potential fix idea)

Meanwhile, we are near to drop 32bits packages (and CI). LLVM 8 is not even available in their apt repo for 32 bits.

Before dropping 32 bits CI (maybe temporarily) this PR introduce a couple of changes and improvements in the CI.

• For 32 bits only std_specs are run
• On nightlies, maintenance, and tagged releases the std_specs are run using the recently built compiler.
  • this recently built compiler fails less often, but there build process is different from the one used to run the per-commit workflow.
• The distributions jobs can start before the test jobs finished.
  • Darwin builds usually took longer to start, delaying the build queue.
  • Failures along the way might cause longer delays.
• How build output is stored in artifacts also changed a bit
  • docker images dump is no longer persisted (it was never actually used)
  • steps that produce files for the final artifacts are also available in each independent steps.

This PR is marked as draft until the changes in distributions-scripts are merged in master.
A sample of the final workflow is already available at https://circleci.com/workflow-run/2fbeef10-7614-4805-aedf-702ab979477b