percpu: speed up virtual address calculations

The percpu address calculations are something like

  addr = base + (var - start)

where

  base is the per-cpu base of the per-cpu area
  var is the variable address in the per-cpu template in the elf image
  start is the beginning of the per-cpu template in the elf image

This patch simplifies the calculation by subtracting start from base.  This
allows us to compute addr with just one addition.