Don't use f32::round for filter 1D

[RunDevelopment]

Same as #2837. I removed f32::round and replaced it with + 0.5 for better perf on x86..

The self.min(u16::MAX as f32) might be unnecessary too. I left it in for now, because it maps NaN to u16::MAX instead of 0 (by the as cast). If NaN isn't a concern for filter 1D, then it can be removed too. Just let me know.

[197g]

I'm actually not sure myself..

[awxkee]

NaNs typically saturated to the lower bound if we assume ieee standard. However, NaNs appearing here is unlikely so I don’t think it’s important.

[197g]

We might put it into compatibility notes but we don't provide a method that would filter float images but convert the result into a uint image. I don't think you could provoke a NaN-conversion from the image input except by the kernel; but then it applies to the whole image which does not seem so useful.

[awxkee]

I think we've overlooked this a bit in those PRs, since I already put filter on not normal values. However, it's still possible to submit here NaN values. I don't see much point to mess with NaN inputs, if you see that different we could just return for any NaN input a default image filled with pre-defined value. exp is defined for all inputs except NaN so in practice except a system has malformed libm here should not be NaNs. As well we could scan kernel and if there is a NaN somewhere it's also possible to return a default image.
I could do PR with NaNs filter or return default image, or something else?