- Win32: when using the threaded RTS, bypass the Async IO stuff and
  just make blocking calls to read()/write().

  This gives a significant performance boost to programs doing lots of
  multithreaded I/O: in fact, a test program I have which does I/O
  over 500 pipes simultaneously goes twice as fast with this change,
  and is even faster than the non-threaded RTS (Windows only - Unix
  changes are in the pipeline too).

- openFd: take an extra parameter to specify socketness of the file
  descriptor rather than assuming that all streams are sockets.  Some
  streams (eg. pipes) aren't sockets.