Should we somehow expose Pipe*Stream on Unix, and if so, how?

[njsmith]

We have PipeReceiveStream and PipeSendStream classes on Unix, that are currently not exposed publicly; they're only used internally for subprocesses.

The actual code is a pretty generic building block: they can wrap around any Unix file-descriptor that supports non-blocking mode. So they could potentially be useful for talking to FIFOs, TTYs, stdio streams (#174), etc.

Should we expose them more directly? If so, how? In trio.hazmat, with some sort of os.fdopen equivalent?

If we do we'll want to think through the ownership semantics: should they take ownership of a passed in fd, or should they dup it?

Also, they should restore the fd's blocking/non-blocking state before closing it.