- Fix more places where custom header size was calculated exclusive of
  the data size member.
- Optimize custom marshaling by getting size before calling the custom
  marshaler so we can write the header before and not use a second
  stream.