Remove bogus check that larger floating point types have smaller minimum

normalized exponents. That's not true for double versus double double.

llvm-svn: 260610