Fix float to LONGEST conversion.

Fix float to LONGEST conversion.

The code used to have undefined behaviour when template parameter is float and
host_float is NaN, because it attempted to convert NaN value to LONGEST at the
last statement. This frequently caused crashes on tests that checked "info
all-registers" (at least when the code is compiled with clang; I didn't test
with gdb).

gdb/ChangeLog:

	*target-float.c (host_float_ops<T>::to_longest): Update
	implementation.