More cmp.c shrinkage.

More cmp.c shrinkage.

Abuse loopfiles() to handle both arguments, caching the first pass in globals.
That eliminates get_fd() since loopfiles already knows how to handle "-" arg.
Chop toybuf in half and use half for each file, eliminating malloc/free.
main() becomes just a call to loopfiles(), which does all open/close.