CI: skip ninja installation in linux_qemu workflows

[mayeut]

The ninja used in the workflow is the one from the host.
Skipping ninja installation in the container allows to workaround issues that could arise when building it from source as is currently the case with riscv64.

xref: scikit-build/ninja-python-distributions#226 (comment)
cc @mattip

[charris]

Could add a comment about why the ninja requirement is removed. I'm wondering what upstream change caused the sudden appearance of these failures.

[mattip]

there really is no good reason to build ninja from source inside the container. So this PR is desirable even if we could have built ninja successfully.

[mayeut]

I'm wondering what upstream change caused the sudden appearance of these failures.

The build-backend changed from scikit-build to scikit-build-core. There are no c++ modifications. I don't know yet what caused the build to fail here (The build is passing native/qemu riscv64 compilation. NumPy is using a cross-compilation setup which I've not tried to reproduce yet).

[mattip]

Thanks @mayeut

[mayeut]

Just FYI, the failure comes from the following:

• scikit-build-core uses sysconfig.get_config_var("CC") when valid and CC not in environment variables, scikit-build wasn't
• NumPy cross-compilation overrides the gcc executable/symlink but not the riscv64-linux-gnu-gcc one which is returned by sysconfig.get_config_var("CC"). It also removes / overrides a couple directories from container to use the host ones which most likely leaves the native/container compiler in a bad state and/or the proper dev dependencies were not there in the first place for the native compiler.

Those 2 factors explain why the compilation is now failing for ninja (CC=gcc CXX=g++ python -m pip install ninja works).

[mattip]

@mayeut thanks for digging into this.