I've committed this to gomp4 branch.  It changes the oacc_xform_on_device and 
oacc_xform_dim_pos calls to lose the GSI parameter.  Thus they look similar to 
the interface we're implementing for the reduction transformations.  The 
iterator in the caller is advanced before the call, so it isn't left holding  a 
potentially deleted instruction.

I noticed a  missing break which would have caused us to fall into the fork/join 
handling from the lock/unlock handling.

nathan