Math: work around a MSVC 2017 15.8 ICE.

Four hours of my life, gone. Minimal repro case here:
https://twitter.com/czmosra/status/1038610648568147968

doc/changelog.dox

@@ -311,6 +311,14 @@ See also:
 
 @subsection changelog-latest-bugfixes Bug fixes
 
+-   MSVC 15.8 (released on Aug 14, 2018) has a regression causing the compiler
+    to crash with an ICE (C1001) on @ref Math::Color4 constructors that have a
+    default alpha parameter. This is only only when the `/permissive-` flag is
+    specified (which is done by default for projects created directly using
+    Visual Studio but not for projects using CMake). This has been worked
+    around, but can't be guaranteed it was the only case. In case you're
+    experiencing a similar issue somewhere else, please try to disable the
+    `/permissive-` flag first.
 -   An assert was firing during @ref Platform::GlfwApplication initialization
     due to a pointer not being properly initialized after the
     @ref Platform::GlfwApplication::GLConfiguration "GLConfiguration" rework in

src/Magnum/Math/Color.h

@@ -197,12 +197,26 @@ template<class T> inline Vector3<typename Color3<T>::FloatingPointType> toXyz(ty
 }
 
 /* Value for full channel (1.0f for floats, 255 for unsigned byte) */
+#if !defined(CORRADE_MSVC2017_COMPATIBILITY) || defined(CORRADE_MSVC2015_COMPATIBILITY)
+/* MSVC 2017 since 15.8 crashes with the following at a constructor line that
+   calls this function via a default parameter. This happens only when the
+   /permissive- (yes, there's a dash at the end) flag is specified, which is
+   projects created directly using VS (enabled by default since 15.5) but not
+   projects using CMake. Not using SFINAE in this case makes it work. Minimal
+   repro case here: https://twitter.com/czmosra/status/1038610648568147968 */
 template<class T> constexpr typename std::enable_if<std::is_floating_point<T>::value, T>::type fullChannel() {
     return T(1);
 }
 template<class T> constexpr typename std::enable_if<std::is_integral<T>::value, T>::type fullChannel() {
     return Implementation::bitMax<T>();
 }
+#else
+template<class T> constexpr T fullChannel() { return bitMax<T>(); }
+/** @todo half */
+template<> constexpr float fullChannel<float>() { return 1.0f; }
+template<> constexpr double fullChannel<double>() { return 1.0; }
+template<> constexpr long double fullChannel<long double>() { return 1.0l; }
+#endif
 
 }