Split -enable-finite-only-fp-math to two options:

-enable-no-nans-fp-math and -enable-no-infs-fp-math. All of the current codegen fp math optimizations only care whether the fp arithmetics arguments and results can never be NaN.

llvm-svn: 108465