Partly mitigate bad Clang inlining decision

Because a visitor is wrapped several times during visitation it cases extra
temporaries usage and useless store and loads that can only be optimized if
the `visitation_impl` is inlined into the function that creates the wrapper.
Clang inliner decides not to inline functions even with small-sized switches,
resulting in a poor visitation code. Forceinline mark on those internal functions perceptibly improves the situation, though does not mitigate it
completely.

LLVM ticket https://bugs.llvm.org/show_bug.cgi?id=41491

include/boost/variant/detail/visitation_impl.hpp

@@ -190,7 +190,7 @@ template <
     , typename Visitor, typename VoidPtrCV
     , typename NoBackupFlag
     >
-inline typename Visitor::result_type
+BOOST_FORCEINLINE typename Visitor::result_type
 visitation_impl(
       const int internal_which, const int logical_which
     , Visitor& visitor, VoidPtrCV storage

include/boost/variant/variant.hpp

@@ -2355,7 +2355,7 @@ class variant
 #endif// !defined(BOOST_NO_MEMBER_TEMPLATE_FRIENDS)
 
     template <typename Visitor, typename VoidPtrCV>
-    static typename Visitor::result_type
+    BOOST_FORCEINLINE static typename Visitor::result_type
     internal_apply_visitor_impl(
           int internal_which
         , int logical_which
@@ -2380,7 +2380,7 @@ class variant
     }
 
     template <typename Visitor>
-    typename Visitor::result_type
+    BOOST_FORCEINLINE typename Visitor::result_type
     internal_apply_visitor(Visitor& visitor)
     {
         return internal_apply_visitor_impl(
@@ -2389,7 +2389,7 @@ class variant
     }
 
     template <typename Visitor>
-    typename Visitor::result_type
+    BOOST_FORCEINLINE typename Visitor::result_type
     internal_apply_visitor(Visitor& visitor) const
     {
         return internal_apply_visitor_impl(