Alternate, simpler PR after feedback on #10398. We can iterate on this one but possibly fall back to the other one if this doesn't work out.

In my testing, installing from a .tar.gz appears to work (i.e. I can import the sklearn module afterwards) but there are some errors displayed during installation. I don't know if those are a concern or not. Full output below.

Installing from a wheel worked smoothly.

Installing from source tarball (Mac OS X Sierra, Python 2.7.13):

$ pip install scikit-learn/dist/scikit-learn-0.20.dev0.tar.gz
Processing ./scikit-learn/dist/scikit-learn-0.20.dev0.tar.gz
Collecting numpy>=1.8.2 (from scikit-learn==0.20.dev0)
  Downloading https://artifactory.rsglab.com/artifactory/api/pypi/pypi/packages/ea/df/f0671353e3d2eab4c87df014ad09505268561f6b09d0dc257d1b32653cfe/numpy-1.13.3-cp27-cp27m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (4.6MB)
Collecting scipy>=0.13.3 (from scikit-learn==0.20.dev0)
  Downloading https://artifactory.rsglab.com/artifactory/api/pypi/pypi/packages/4d/e4/e92135b070c0913cbee59849b61f57076ac33d8a754be0fef581a28676f9/scipy-1.0.0-cp27-cp27m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (16.8MB)
Building wheels for collected packages: scikit-learn
  Running setup.py bdist_wheel for scikit-learn: started
  Running setup.py bdist_wheel for scikit-learn: finished with status 'error'
  Complete output from command /Users/bhart/.pyenv/versions/2.7.13/envs/scikit-learn-2.7.13/bin/python2.7 -u -c "import setuptools, tokenize;__file__='/private/var/folders/vd/j7yv3l2x0tb2t_7pvwrvhr2r0000gq/T/pip-GZj9kO-build/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\r\n', '\n');f.close();exec(compile(code, __file__, 'exec'))" bdist_wheel -d /var/folders/vd/j7yv3l2x0tb2t_7pvwrvhr2r0000gq/T/tmp5WQUMIpip-wheel- --python-tag cp27:
  Partial import of sklearn during the build process.
  Traceback (most recent call last):
    File "/private/var/folders/vd/j7yv3l2x0tb2t_7pvwrvhr2r0000gq/T/pip-GZj9kO-build/setup.py", line 149, in get_numpy_status
      import numpy
  ImportError: No module named numpy
  Traceback (most recent call last):
    File "<string>", line 1, in <module>
    File "/private/var/folders/vd/j7yv3l2x0tb2t_7pvwrvhr2r0000gq/T/pip-GZj9kO-build/setup.py", line 241, in <module>
      setup_package()
    File "/private/var/folders/vd/j7yv3l2x0tb2t_7pvwrvhr2r0000gq/T/pip-GZj9kO-build/setup.py", line 231, in setup_package
      .format(numpy_req_str, instructions))
  ImportError: Numerical Python (NumPy) is not installed.
  scikit-learn requires NumPy >= 1.8.2.
  Installation instructions are available on the scikit-learn website: http://scikit-learn.org/stable/install.html
  
  
  ----------------------------------------
  Running setup.py clean for scikit-learn
Failed to build scikit-learn
Installing collected packages: numpy, scipy, scikit-learn
  Running setup.py install for scikit-learn: started
    Running setup.py install for scikit-learn: still running...
    Running setup.py install for scikit-learn: finished with status 'done'
Successfully installed numpy-1.13.3 scikit-learn-0.20.dev0 scipy-1.0.0
$ python
Python 2.7.13 (default, Jun  8 2017, 17:00:51)
[GCC 4.2.1 Compatible Apple LLVM 8.0.0 (clang-800.0.42.1)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import sklearn
>>> import numpy
>>> import scipy
>>>

Fixes #10398, fix #7867.