Elena, this splits off two versions of push_dummy_call: one for sh2e,
sh3e, and sh4, and the other for all the others.  There are two 
differences:
how they pass floats, and how they pass large structs.

This wins us 6 new passes in call-ar-st, one in call-rt-st, 13 in 
callfuncs,
and one in varargs.