Switch to ondisk also when unpacking FS

From benchmarks it seems to be still faster. Add a note for a future
improvement

pkg/api/core/types/artifact/artifact_test.go

@@ -156,7 +156,7 @@ RUN echo bar > /test2`))
 			Expect(err).ToNot(HaveOccurred())
 			defer os.RemoveAll(result) // clean up
 
-			img, err := b.ImageReference(resultingImage, false)
+			img, err := b.ImageReference(resultingImage, true)
 			Expect(err).ToNot(HaveOccurred())
 			_, _, err = image.ExtractTo(
 				ctx,

pkg/compiler/backend/simpledocker.go

@@ -140,6 +140,11 @@ func (s *SimpleDocker) Push(opts Options) error {
 }
 
 func (s *SimpleDocker) ImageReference(a string, ondisk bool) (v1.Image, error) {
+	// TODO: We could also handle this from docker's pipe, but needs benchmarking:
+	// daemon.Image takes a client optionally. Otherwise we can return a new image
+	// from a tarball by providing ourselves a reader from docker stdout pipe.
+	// See daemon.Image implementation below for an example (which returns the tarball stream
+	// from the HTTP api endpoint instead ).
 	if ondisk {
 		f, err := s.ctx.Config.GetSystem().TempFile("snapshot")
 		if err != nil {

pkg/compiler/compiler.go

@@ -234,7 +234,7 @@ func (cs *LuetCompiler) unpackFs(concurrency int, keepPermissions bool, p *compi
 		}
 	}
 
-	img, err := cs.Backend.ImageReference(runnerOpts.ImageName, false)
+	img, err := cs.Backend.ImageReference(runnerOpts.ImageName, true)
 	if err != nil {
 		return nil, err
 	}