Better support for Random123 routines in verbatim blocks for GPU

[pramodk]

Consider following example in Ampa and Gaba synapses of BBP model :

VERBATIM
static void bbcore_read(double* x, int* d, int* xx, int* offset, _threadargsproto_) {
  assert(!_p_rng);
  uint32_t* di = ((uint32_t*)d) + *offset;
  if (di[0] != 0 || di[1] != 0 || di[2] != 0) {
      nrnran123_State** pv = (nrnran123_State**)(&_p_rng);
      *pv = nrnran123_newstream3(di[0], di[1], di[2]);
      char which = (char)di[4];
      nrnran123_setseq(*pv, di[3], which);
  }
  *offset += 5;
}
ENDVERBATIM

INITIAL {
        LOCAL tp_AMPA, tp_NMDA
        tp_AMPA = (tau_r_AMPA*tau_d_AMPA)/(tau_d_AMPA-tau_r_AMPA)*log(tau_d_AMPA/tau_r_AMPA)
       ....

        VERBATIM
        if( usingR123 ) {
            nrnran123_setseq((nrnran123_State*)_p_rng, 0, 0);
        }
        ENDVERBATIM
}

As these mod files are compiled for GPU, nrnran123_setseq get replaced with cu_nrnran123_setseq (because of macro in nrnran123.h).

nrnran123_setseq in INITIAL block is called on GPU so its ok to call cu_nrnran123_setseq. But bbcore_read is called in CPU context and calling CUDA kernel cu_nrnran123_setseq result in abort.

[alkino]

I think this is fixed by #1125 and no more writing directly random functions