Box DiagnosticBuilder.

It's a large type -- 176 bytes on 64-bit. And it's passed around and
returned from a lot of functions, including within PResult.

This commit boxes it, which reduces memory traffic. In particular,
`PResult` shrinks to 16 bytes in the best case; this reduces instruction
counts by up to 2% on various workloads.

src/librustc_errors/annotate_snippet_emitter_writer.rs

@@ -37,7 +37,7 @@ impl Emitter for AnnotateSnippetEmitterWriter {
                                           &mut primary_span,
                                           &mut children,
                                           &db.level,
-                                          db.handler.flags.external_macro_backtrace);
+                                          db.handler().flags.external_macro_backtrace);
 
         self.emit_messages_default(&db.level,
                                    db.message(),