Fix GDI leak in NativeThemeWin::PaintIndirect

ui/native_theme/native_theme_win.cc

Index: ui/native_theme/native_theme_win.cc
diff --git a/ui/native_theme/native_theme_win.cc b/ui/native_theme/native_theme_win.cc
index 043a3d47278a278e01d1c455157a9ae06ceec4b9..684ff3010ca561c09df6093bbd87b914d873e6ae 100644
--- a/ui/native_theme/native_theme_win.cc
+++ b/ui/native_theme/native_theme_win.cc
@@ -664,9 +664,11 @@ void NativeThemeWin::PaintIndirect(SkCanvas* destination_canvas,
   //                  be sped up by doing it only once per part/state and
   //                  keeping a cache of the resulting bitmaps.
 
+  // Bitmap needs to be deleted together with DC (before or after).

[Peter Kasting, 2016/09/23 18:45:00]

According to docs I found on the web, an attempt to free the HBITMAP while it's
still selected into a valid DC will fail.  So the destruction order must be DC
first, then bitmap (unless we were to SelectObject() a new HBITMAP into the DC
first).

This seems to contradict the comment here.  The code here happens to get this in
the right order, but I'm worried about future misuse.

Maybe we should nuke CreateOffscreenSurface() and do it inline here.  That way
we can be very clear in the comments about the declaration order of different
things, and we won't risk others getting this wrong.

You'll want to rebase anyway to pick up the changes from
https://codereview.chromium.org/2355423002/ .

+  base::win::ScopedBitmap bitmap;
   // Create an offscreen canvas that is backed by an HDC.
   base::win::ScopedCreateDC offscreen_hdc(
-      skia::CreateOffscreenSurface(rect.width(), rect.height()));
+      skia::CreateOffscreenSurface(&bitmap, rect.width(), rect.height()));
   sk_sp<SkSurface> offscreen_surface =
       skia::MapPlatformSurface(offscreen_hdc.Get());
   DCHECK(offscreen_surface);