Constant fold int-to-long-double conversions;

use APFloat for int-to-float/double; use
round-to-nearest for these (implementation-defined,
seems to match gcc).

llvm-svn: 42484