Scheduling:

 - remove loop/get/chain from GstElement and add a "iterate" method.
   The iterate method is called with the event (or events) that
   triggered it, performs some action, and resets the events (file
   descriptors becoming readable, semaphores, pads becoming readable
   or writable, or a time occurs).

 - Add GstLoopElement, GstChainElement, etc. for compatibility.

 - Remove existing state handling and create 2 states, "playing" and
   "stopped".  "playing" means that the iterate() method of the
   element may be called, that is, the element is allowed to move
   buffers, negotiate, etc.  "stopped" means that no gstreamer-ish
   things happen to an element, only gobject-ish.  A separate
   reset() method will handle the difference between READY and NULL.

 - Add a flag "ready" to GstElement that is under the control of the
   element.  If the element is ready to stream, it sets this flag,
   and the entire pipeline starts streaming.  (This is basically
   the difference between PAUSED and PLAYING.)  For example, osssink
   won't set the ready flag until the device is opened and there is
   a buffer available to write to the device.

 - Scheduling of elements and movement of buffers will be timed by
   clocks.  



Example #1:

 Pipeline: sinesrc ! osssink

 - The application creates the pipeline and sets it to "playing".

 - The clock is created and set to "paused".

 - sinesrc.iterate() decides to watch for the event "src pad
   negotiation" and sets the available caps on the pad.

 - osssink.iterate() opens device, determines available caps, and
   sets the available caps on the pad.  Then it decides to wait for
   "sink pad negotiation".

 - The scheduler realizes that the two elements are waiting for
   negotiation, so it negotiates the link.

 - sinesrc.iterate() sets the "ready" flag (because it needs no more
   preparation to stream) and decides to watch for the event "src
   pad ready to accept buffer".

 - osssink.iterate() decides to watch for the event "sink pad has
   available buffer".

 - The scheduler realizes that sinesrc.srcpad is now ready, so it
   calls sinesrc.iterate()

 - sinesrc.iterate() creates a buffer and pushes it, and decides to
   wait for the same event.

 - The scheduler realizes that osssink.sinkpad now has a buffer, so
   it calls osssink.iterate().

 - osssink.iterate() is now ready to stream, so it sets the "ready"
   flag and waits for "time 0".

 - The pipeline is now completely ready, so the clock may be
   started.  A signal is fired to let the application know this
   (and possibly change the default behavior).

 - The clock starts with the time 0.  The scheduler realizes this,
   and decides to schedule osssink.

 - osssink.iterate() is called, and writes the buffer to the device.
   This starts the clock counting.  (Actually, the buffer could be
   written by the clock code, since presumably the clock is related
   to osssink.)  iterate() then waits for "sink pad has available
   buffer".

 We're now basically in streaming mode.  A streaming cycle:

 - osssink.iterate() decides the audio output buffer is full enough,
   so it waits for "time X", where X is the time when the output
   buffer will be below some threshold.

 - osssink.iterate() waits for "sink pad has available buffer"

 - sinesrc.iterate() creates and pushes a buffer, then waits for
   "src pad ready".


 Further ideas:

 - osssink can set a hard deadline time, which means that if it is
   not scheduled before that time, you'll get a skip.  Skipping
   involves setting osssink to "not ready" and pauses the clock.
   Then the scheduler needs to go through the same process as above
   to start the clock.

 - As a shortcut, osssink can say "I need a buffer on the sinkpad
   at time X".  This information can be passed upstream, and be used
   in filters -- filter.sinkpad says "I need a buffer at time X-N",
   where N is the latency of the filter.


Example #2:

 Pipeline: osssrc ! osssink
 
 - The application creates the pipeline and sets it to "playing".

 - The clock is created and set to "paused".

 - negotiation happens roughly as in example #1, although osssrc
   additionally opens and prepares the device.

 - osssrc.iterate() sets the "ready" flag (because it needs no more
   preparation to stream) and waits for "time 0", since it presumably
   can't wait for the file descriptor (audio input hasn't been
   enabled on the device yet.)

 - osssink.iterate() decides to watch for the event "sink pad has
   available buffer".

 - The scheduler realizes the deadlock and (somehow) tells osssink
   that it can't pre-roll.  (This needs more work)  In other words,
   osssink can't be the clock master, but only a clock slave.

 - osssink.iterates() agrees to start at time SOME_LATENCY, sets the
   "ready" flag, and waits for a buffer on its sink pad.

 - The pipeline is now completely ready, so the clock may be
   started.  A signal is fired to let the application know this
   (and possibly change the default behavior).

 - The clock starting causes two things to happen: osssrc starts
   the recording of data, and osssink starts the outputting of data.
   The data being output is a chunk of silence equal to SOME_LATENCY.

 - osssrc.iterate() is called for "time 0", does nothing, and waits
   on the file descriptor (via the scheduler, of course).  All waiting
   on file descriptors should have an associated timeout.

 - osssrc.iterate() is called when the file descriptor is ready,
   reads a chunk of data, and pushes the buffer.  It then waits for
   its file descriptor to be ready.

 - osssink.iterate() is called


Evil:

  fakesrc ! tee ! fakesink tee0. ! never_accept_a_buffer_sink

  sinesrc ! osssink videotestsrc ! ximagesink
  
  fakesrc ! fakesink (pausing)

  sinesrc ! identity ! osssink