* Convert to use t_vertex.c instead of sis_vb.[ch]

* Convert to use t_vertex.c instead of sis_vb.[ch]
* Don't dispatch vertices directly to MMIO; queue them up in dma-like buffers
  first.  This makes things more uniform between AGP and MMIO paths, cleans up
  some locking ugliness, and makes the driver look more like other drivers.
* Don't use the AGP Cmd buffer provided by the server.  Instead allocate one in
  the client, which avoids the need for lots of synchronization stuff.
* Mark some MMIO accesses volatile that should have been.
* Disable the AGP submission path by default (agp_disable=true) due to
  unresolved issues in the new code.  The old code had its own (serious) errors
  with AGP, so this is not really a step backwards.