docker,ci,make: add Ceph-Dashboard support and other improvements

[epuertat]

This PR is to make it easier for developers to test Ceph-Dashboard when doing integration testing.
Fixes: #357
Fixes: #382

docker: add Ceph-Dashboard support

Include ceph-mgr-dashboard package and deps (grpcio).

Create new ceph-devel service to mount local ceph repo inside the container (for dev purposes).

docker: add ceph-devel container

Also:

• Extend healthcheck retries to prevent random failures
• Configure nvmeof gateway in Ceph
• Fix nvmeof-cli aliases (the leading space prevented the shell from
  adding these to the history)
• Add syntax highlight tag to the Dockerfile.spdk

docker: add custom Ceph repo

• A custom Ceph YUM repo (usually a ceph-ci generated/Chacra one, like
  https://3.chacra.ceph.com/r/ceph/ceph-nvmeof-mon/c1ced49d133c2bcd1fa9311a35688bf641b93c76/centos/9/flavors/default/)
• New make targer clean_cache to prune builder cache (required when
  changing the Ceph YUM Repo.
• Repo/RPM GPG check disabled. Not ideal, but ceph-ci builds don't sign
  RPMs.
• Specify ulimits (nofile 1024) for the Ceph container. In Docker CE,
  ceph-mon's ms_dispatch thread iterated through all available file
  descriptors and tried to close them. By setting nofile to 1024, this
  is heavily constrained. This behaviour is not exhibited with
  Fedora's Moby-engine.

docker: support target-arch in SPDK

• Remove exising handling for AVX512*, VPCLMULQDQ and RDSEED
  instructions.
• Replace it with spdk/configure --target-arc=
• Tune SPDK's yum/dnf cache and settings.

ci: remove nvmeof-devel dependency

nvmeof-devel is not required for pytests, since they're mounted via run CLI command.

make: fix alias issue

Alias was not properly working for cephnvmf* as it was erased from the shell history.

docker: save SPDK target arch as label

SPDK/DPDK relies on pretty modern CPU instruction sets, which renders that incompatible with certain servers (Ivy Bridge and older). This new SPDK_TARGET_ARCH setting allows setting that. The new default (x86-64-v2) could result in a less performant SPDK.

ci: update upload-download artifact actions

The v4 of these actions provide 90% speed-ups by compressing artifacts.
Migration to this new action means zero cost.

Additional core clean-up performed.

ci: split ceph job and update checkout

There's no need for ceph to be built together with the rest of the nvmeof services. In fact this container could/should be pulled off to a separate repo/CI.

ci: cancel CI run when another starts

When a user pushes a change to a PR, if there was an ongoing CI run,
this will cancel any previous runs. This will save some CO2.

[baum]

lgtm 🖖

[epuertat]

@gbregman please review this PR as it removes part of the fixes you introduced for the missing CPU instruction.

@baum I pulled off the --target-arch configure param to a separate env. var, in order to make it easier to generate multiple SPDK versions from matrix job (e.g.: x86-64-v2 and x86-64-v4 might be enough to cover legacy and modern CPUs). Probably not required in every build, but just when publishing a release.

[gbregman]

LGTM, only I don't really know much about these files. One comment though, if you removed the SPDK_DISABLE_VPCLMULQDQ and SPDK_DISABLE_AVX512 vars, shouldn't you also remove them from .github/workflows/build-container.yml?

[gbregman]

@epuertat notice that on top of disabling the AVX512 and the MULQDQ instructions we also disable RDSEED. Only this id hardcoded in Dockerfile.spdk:

# Disable RDSEED, see https://github.com/ceph/ceph-nvmeof/issues/259
RUN \
    sed -i "s/^\( \+'RDSEED'\)/#\1/" dpdk/config/x86/meson.build