Factor the constraint-favoring machinery out of the constraint generation process, and re-work it into a series of passes over an expression sub-tree.

Aside from tidying things up, doing this results in some significant benefits:
- Allows for global constraint ordering optimizations over a given expression, not just on a peephole basis.
- Eliminates a set of order-dependent bugs in the solver that have been dogging us for a while. (rdar://problem/19459079)
- Brings another set of tyvar-to-tyvar solving problems out of the realm of the exponential. (rdar://problem/19005271)
- Opens up the possibility of optimizing constraints during later solving phases - not just while generating them.

Swift SVN r24693

stdlib/core/Unicode.swift

@@ -867,7 +867,7 @@ extension UTF16 {
   /// Requires: `width(x) == 2`
   public static func leadSurrogate(x: UnicodeScalar) -> UTF16.CodeUnit {
     _precondition(width(x) == 2)
-    return UTF16.CodeUnit((x.value - 0x1_0000) >> 10) + 0xD800
+    return UTF16.CodeUnit((x.value - 0x1_0000) >> (10 as UInt32)) + 0xD800
   }
 
   /// Return the low surrogate code unit of a `surrogate pair