De-pollute parameters to bdist so its only on osx and make sure that osx also sets the correct numpy version for python3.7

This reverts commit 47c5a9fb.

xcode7.3 image has too old clang which does not support thread-local storage -> update to xcode8 image

* Adjust build environment as per https://docs.opencv.org/3.1.0/d5/de5/tutorial_py_setup_in_windows.html and https://wiki.python.org/moin/WindowsCompilers#Which_Microsoft_Visual_C.2B-.2B-_compiler_to_use_with_a_specific_Python_version_.3F

* The best practice is to only use requirements.txt for deployments, not for packaging

* ignore PyCharm metadata

* refactor to be more modular

* the warnings are irrelevant in older versions

* v110 toolset might introduce dependency on older runtime not likely to be present in newer systems, potentially breaking upgrade

* build locally with scikit-build

* adjust appveyor config to local build

* cleanup

* DLL load works without

* fail appveyor build immediately when one job fails

* * cmd doesn't expand globs
* - diplication
* - WinXP-related bits
* pip erroneously considers cv2/ as installed package when in sourceroot
* require CMake output entries to be found
* 3.5+ uses different .pyd naming scheme
* .pyd is placed differently in Linux
* DLL is only in Windows

* ignore temporary config files

* Yet another .pyd naming convention in Py35 Linux.
Screw checking.

* Update setup.py file to flag compatibility with Python 2.7-3.4-3.5-3.6 (#57)

# Conflicts:
#	setup.py

* Adjust build environment as per https://docs.opencv.org/3.1.0/d5/de5/tutorial_py_setup_in_windows.html and https://wiki.python.org/moin/WindowsCompilers#Which_Microsoft_Visual_C.2B-.2B-_compiler_to_use_with_a_specific_Python_version_.3F

* The best practice is to only use requirements.txt for deployments, not for packaging

* ignore PyCharm metadata

* refactor to be more modular

* the warnings are irrelevant in older versions

* v110 toolset might introduce dependency on older runtime not likely to be present in newer systems, potentially breaking upgrade

* build locally with scikit-build

* adjust appveyor config to local build

* cleanup

* DLL load works without

* fail appveyor build immediately when one job fails

* * cmd doesn't expand globs
* - diplication
* - WinXP-related bits
* pip erroneously considers cv2/ as installed package when in sourceroot
* require CMake output entries to be found
* 3.5+ uses different .pyd naming scheme
* .pyd is placed differently in Linux
* DLL is only in Windows

* ignore temporary config files

* fail fast on error

* * custom build logic should be unneeded now, will uncomment as needed

* use the latest upstream multibuild

* enable tracing
- more suspicious things

* save some build time

* these somehow cause "Server does not allow request for unadvertised object" for GitHub

* Revert "use the latest upstream multibuild"

This reverts commit 915ba421f97acebb3edcf55b5e46ea7aab8fea42.

* Yet another .pyd naming convention in Py35 Linux.
Screw checking.

* Update setup.py file to flag compatibility with Python 2.7-3.4-3.5-3.6 (#57)

# Conflicts:
#	setup.py

* add diagnostics

* * Use upstream multubuild due to the forked one abusing prebuild as build

* save some time

* looks like clean_code() isn't required

* * Don't make CMake try Ninja first
* MacOS LAPACK on Travis is still broken

* See build progress as it goes

* update submodules on demand

* fetch source in time for find_version.py

* use custom docker image

* use custom multibuild

* bump submodule

* Linux/OSX are supposed to build with Qt support

* * make shallow copies to save on download time
* manylinux git is an old version

* * clean up redundant syntax
* reformat codeas YML multiline with dedent

* IPP IW broken in Linux

* + ffmpeg patch

* additional deps required due to multibuild bug

* +cv2.data

* Use upstream multibuild with monkey patch

* don't do anything if upload not needed

* filter out irrelevant debug output

* * clean up unused scripts
* inline the remaining standalone scripts

* https://github.com/matthew-brett/multibuild/issues/106 fixed

* * bump OpenCV version to 3.4.0

* * https://github.com/opencv/opencv/pull/10011 is available in 3.4.0
* allow for if the sourcetree is a submodule

* * don't reinstall whatever happens to be preinstalled
* re-add workaround for https://github.com/travis-ci/travis-ci/issues/9055

* * rearrange custom CMake flags
* -DWITH_GTK=OFF proved to be unneeded, gthread is among manylinux requirements

* fix linux detection to work in 3.3+, too

* fix data path for other OSes

* don't do anything unless deployment is requested

* use upstream ``devel`` multibuild

* restore from backup after rebase

* catch fetch error

* use `devel` multibuild

* comment not using --depth

* remove redundant command after merge

automatically on macOS

* Moved license files to package_data

* Moved license files

* Copying licenses to cv2 in Travis step

* Added license copy to Windows build script

* Relax numpy version requirements

* trying to make apt-get happy

* Remove requirements.txt installation

This installs numpy, which now comes from travis.yml

* Move numpy's dependency to appveyer

* Install numpy on OSX

* Drop numpy required version to 1.11.1 for Pythons that aren't 3.6

* Make sure that the test depends match the build depends

This reverts commit 3396bdf7.