This fixes ELF-header generation in 0l/vl so that the early
parts of the header are byte-order-independent and the later
parts are in the target byte order.  I believe the alignment
claims made by the header are more accurate, too.

This has been tested for little-endian targets only, but produces
ELFs which are loadable by the Simics system simulator and which
"make sense" to readelf and objdump.