The test takes the following steps using two threads. The threads are
only used so it can operate on another context without having to
rebind it. The threads are run lock-step so that each step is run
sequentially.

Thread 1: Make a flushless context A
Thread 1: Make a flushy context B, shared with A
Thread 1: Make a flushy context C, shared with A
Thread 1: Bind context A
Thread 2: Bind context C
Thread 1: Make a renderbuffer.
Thread 1: glClear() it to green.
Thread 1: glFinish()
Thread 1: glClear() it to red.
Thread 2: Do a glReadPixels()

(At this point the GL implementation is allowed to have finished
the clear to red but it probably won't have. If the read pixels
returns green here then it's not a failure but the test won't work
so it will report PIGLIT_SKIP)

Thread 1: Bind context C
Thread 1: sleep(.5)
Thread 2: Make sure glReadPixels() is still green, otherwise fail.

All of the steps are then run again but this time context A is made
flushy and the last step ensures that the pixel becomes red instead
of green. If it did become red then the GL successfully made a
flush when context A was released.

The test also verifies that calling glGetIntegerv with
GL_CONTEXT_RELEASE_BEHAVIOR returns the expected value when setting
the attribute to none and flush and also when the attribute is left
out entirely.