fix exp(NaN16) and add tests

[oscardssmith]

Fixes #42554

[KristofferC]

The N = unsafe_trunc(Int32, N_float) can be moved down to under this check now, or?

Also, I don't understand how the unsafe_trunc is safe. Are there really only values that fit in Int32 that can possibly reach that call?

[oscardssmith]

It could be moved, but it won't change performance so I won't to keep the flow closer to that of the other implementations (especially since the computation is still needed in the common case). The reason why this is safe is that we are taking exp(x) so if round(x) large exp(x) will overflow, and if it is small, exp(x) will underflow. As such even for Float64, N must be in (-1075,1024). All other values of N will just produce 0 or Inf

[KristofferC]

Rebased to hopefully get a better CI run

[oscardssmith]

Rebased with fixed tests. Lets see if this will finally work.

[oscardssmith]

Ready to merge?

[oscardssmith]

Merging tomorrow if there aren't objections.

[oscardssmith]

what do I need to do to backport this?

[KristofferC]

Nothing, it has the label so it will get backported. In fact, it already is :)