Remove the API compatibility layer which converted add, sub, and mul

to fadd, fsub, and fmul, when used with a floating-point type. LLVM
has supported the new instructions since 2.6, so it's time to get
on board.

llvm-svn: 102971