Reimplement outer rounds as "compact" in x86 assembler. This has rather

strong impact on decrypt performance, 20-25%. One probably should consider
switching between slower and faster routines depending on how much data
we were asked to process.