The situation is that the driver performs a GPU context switch when
changing to different X-servers on different VTs. Means: you can't
have two X-servers simultaneously generate output, only one can
be operational at a time.
Unfortnuatly, there doesn't exist an OOTB solution for what you
like to achive. A somewhat simple solution would be to define
two seperate screens and to start a Xnest-Server (or Xepher)on
screen 0:1. However, Xnest/Xepher would need to be extended to receive
keyboard and mouse input events from external devices rather
than taking forwarded events from the accelerated X-server. This
is probably just a couple of days of implementation effort, but
the drawback is that the Xnest screen wouldn't have 3D acceleration ...
Edit: just run into that one:
Seems that somebody already did the necessary modifications to Xephyr.