GL testing on the CIs

[mosra]

Moving this out of #99 so I can finally, after two years, close it and together with that also around fifteen tabs in my browser.

Currently all the CIs test only code that doesn't execute anything from OpenGL. This makes the test coverage pretty poor and also doesn't uncover accidental breakages (because locally I'm testing mainly on NV cards where the driver bugs are sparse). Possible options are:

• Build llvmpipe Mesa on Travis and use it as a software GL driver (needs whitelisting for quite a bunch of X11 packages which will take ages). I should test this locally first.
• Build SwiftShader on Travis or AppVeyor and use it as a software GL driver. It's a Google code, so I assume ugly buildsystem issues like the ones in Build for ANGLE on AppVeyor #206. Ugh.
• Maybe there is a CI that provides headless GPU virtualization?

Links:

• https://docs.travis-ci.com/user/gui-and-headless-browsers/
• OpenGL 3+ support travis-ci/travis-ci#5645
• APT whitelist request for libosmesa6-dev travis-ci/travis-ci#3524
• https://stackoverflow.com/a/36375772
• http://help.appveyor.com/discussions/questions/2522-running-unit-tests-opengl
• Use OpenGL Mesa in Appveyor tests Wangscape/Wangscape#117

[mosra]

Some early SwiftShader bugs were fixed lately, which should make this doable. Initial plan is to host a SwiftShader binary on ci.magnum.graphics and testing with that one. Should bump code coverage from 51% to over 80%.

[mosra]

Done as of d2d7ee7, mosra/magnum-integration@deb8c45, mosra/magnum-extras@02e9c5d ... and since 834c5fe and mosra/magnum-plugins@e6f8792, no GL testing needs to be done for the plugins as they are all now completely independent on the GPU API used -- which also reduces the CI running time for them to less than 1/3.

🎉 🍻