Fixes a potential crash when encoding NaNs

This fixes tansig_approx() to avoid crashing when the input is NaN.
The problem could only be triggered when calling the float API with
a float build at a complexity of 7 or more (i.e. analysis called).
Since the crash was due to an out-of-bound read (typically the index
is INT_MIN), it's unlikely to be exploitable in any other way than
causing a crash.