Reimplement the test based on code from adg@golang.org.

The previous version has a race since the file is closed via defer
rather than in the go routine. This meant that the file could be
closed before the go routine has actually received io.EOF. It then
receives EBADF and continues to do zero-byte writes to the pipe.

This addresses an issue seen on FreeBSD and OpenBSD, where the test
passes but exits with a SIGPIPE, resulting in a failure.

R=golang-dev, adg
CC=golang-dev
https://golang.org/cl/5554083