Performance slowdown in sample_get_x from the checks for warnings at the end

[valassi]

I am doing a few tests with sample_get_x towards vectorising it, see #963

I realised that there is a relatively large overhead there from some checks for warnings at the end. In a cuda only build of DY+3jets (CMS #943 from @choij1589) this seems to take around 15 to 20% for some subprocesses.

The question is whether these checks must be kept and in which form. I actually realised becase I removed them in my vectorization tests. This type of checks is likely to be difficult to vectorize. I agree some checks are needed, but maybe this could be reviewed.

This is these checks here (by the way there may be a minor bug as icount is not SAVEd)

c
c     Simple checks to see if we got the right point note 1e-3 corresponds
c     to the fact that the grids are required to be separated by 1e-14. Since
c     double precision is about 18 digits, we expect things to agree to
c     3 digit accuracy.
c
      if (abs(ddum(j)-xbin(x,ij))/(ddum(j)+1d-22) .gt. 1e-3) then
         if (icount .lt. 5) then
            write(*,'(a,i4,2e14.6,1e12.4)')
     &           'Warning xbin not returning correct x', ij,
     &           ddum(j),xbin(x,ij),xo
         elseif (icount .eq. 5) then
            write(*,'(a,a)')'Warning xbin still not working well. ',
     &           'Last message this iteration.'
         endif
         icount=icount+1
      endif

@oliviermattelaer any feedback? Thanks Andrea