Make median! type stable for small float types

[andreasnoack]

Fixed #29900

[StefanKarpinski]

I have an alternative fix that returns the first NaN value, preserving the payload of the specific NaN. Not sure if we care about that, but it is how arithmetic operations behave.

[andreasnoack]

I don't have strong feelings here but on my machine there seems to be a time penalty for using findfirst relative to any and all floating point vectors will pay. I'm expecting few people will care much about getting the right kind of NaN when passing an abstractly typed vector. However, I'm fine with your judgement here so feel free to merge the one you prefer.

[StefanKarpinski]

Yeah, probably doesn't matter. Unfortunate that findfirst has a performance penalty but I agree.