[builtins] fix Torque-implemented ToString performance regression

The order of the typeswitch branches causes repeated Smi-checks.
This CL fixes this by putting the Number case first.
However, the generated code is still worse due to repeated Map and
InstanceType loads. This will be fixed by a future load elimination for
Torque/CSA.

Bug: chromium:955976
Change-Id: I0f59ef795878f65b3cb11246626738bc33f8aff5
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1581644
Commit-Queue: Tobias Tebbi <tebbi@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/master@{#60986}

src/builtins/base.tq

@@ -2536,26 +2536,26 @@ transitioning macro ToStringImpl(context: Context, o: Object): String {
   let result: Object = o;
   while (true) {
     typeswitch (result) {
-      case (str: String): {
-        return str;
-      }
       case (num: Number): {
         return NumberToString(num);
       }
+      case (str: String): {
+        return str;
+      }
       case (oddball: Oddball): {
         return oddball.to_string;
       }
-      case (Symbol): {
-        ThrowTypeError(kSymbolToString);
-      }
       case (JSReceiver): {
         result = NonPrimitiveToPrimitive_String(context, result);
         continue;
       }
+      case (Symbol): {
+        ThrowTypeError(kSymbolToString);
+      }
       case (Object): {
         return ToStringRT(context, o);
       }
     }
   }
-  return ToStringRT(context, o);
+  unreachable;
 }