[REVIEW] Avoid building cython twice

[galipremsagar]

Description

This PR moves the build_ext step to only happen when INSTALL_TARGET=='', thus avoiding the cython build to occur twice when build.sh is invoked.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[bdice]

This looks good to me -- but I feel like there was some context about why this works as it does. @vyasr Was there some issue where scikit-build should have known to "just install" in the second command? If not, then this is probably fine to merge as-is.

[ttnghia]

The issue annoyed me for a long time. Thanks for this.

[galipremsagar]

/merge

[vyasr]

This looks good to me -- but I feel like there was some context about why this works as it does. @vyasr Was there some issue where scikit-build should have known to "just install" in the second command? If not, then this is probably fine to merge as-is.

The pair of commands predate the scikit-build transition. At the time I left it as is because I assumed that users of build.sh may originally have been relying on being able to import the package from inside the directory. The main difference between the behavior of build_ext followed by install vs just install (before the pyproject.toml transition) is that when only doing install you'll get import errors when you try to e.g. import cudf from inside the python/cudf directory because Python will prefer the local directory even though it's not a package. Running both commands makes that work as expected. However this is going to become the norm anyway when we move build.sh to use pip install commands, so I'm OK starting to impose that limitation now. Moving metadata to pyproject.toml has other subtle implications that we may as well start figuring out now as well.