[ClangImporter] Use Clang's -fembed-bitcode mode when under -embed-bitcode.

This should have no functionality change, but is supposed to keep us from
accidentally relying on the "full" Clang importer when in a backend job.
I tested it by archiving a little iOS app from a developer.

Unfortunately, part of the motivation here was that we'd get error messages when
we pass something Clang doesn't like, and that doesn't seem to be happening.

rdar://problem/21389553

Swift SVN r30407

lib/Frontend/CompilerInvocation.cpp

@@ -769,6 +769,9 @@ static bool ParseClangImporterArgs(ClangImporterOptions &Opts, ArgList &Args,
 
   Opts.DumpClangDiagnostics |= Args.hasArg(OPT_dump_clang_diagnostics);
 
+  if (Args.hasArg(OPT_embed_bitcode))
+    Opts.Mode = ClangImporterOptions::Modes::EmbedBitcode;
+
   return false;
 }