Use the APInt versions of the bit-wise conversions of float/double to int

and back. While this is not strictly necessary, it does pave the way for
future changes in casting. It should now be possible to improve vector
casting to deal with un-equal fp/int sizes.

llvm-svn: 34808