Always split String#<< to specialize encoding negotiation and TruffleString.ConcatNode per call site

* Encoding negotiation is more efficient that way.
* That way ConcatNode will see much less polymorphism between empty left, empty right, total length < LAZY_CONCAT_MIN_LENGTH, MutableTruffleString, etc.
* Similar reasoning to to Array#<<.

Author: eregon

src/main/java/org/truffleruby/core/string/StringNodes.java

@@ -101,6 +101,7 @@
 import org.jcodings.specific.ASCIIEncoding;
 import org.truffleruby.Layouts;
 import org.truffleruby.annotations.CoreMethod;
+import org.truffleruby.annotations.Split;
 import org.truffleruby.builtins.CoreMethodArrayArgumentsNode;
 import org.truffleruby.annotations.CoreModule;
 import org.truffleruby.annotations.Primitive;
@@ -417,7 +418,7 @@ RubyString dupAsStringInstance(Object string,
         }
     }
 
-    @CoreMethod(names = "<<", required = 1, raiseIfNotMutableSelf = true)
+    @CoreMethod(names = "<<", required = 1, raiseIfNotMutableSelf = true, split = Split.ALWAYS)
     @ImportStatic(StringGuards.class)
     public abstract static class StringConcatOneNode extends CoreMethodArrayArgumentsNode {