Recognize NaN operands in Min and Max ops (#19984)
### Description
Update the Min and Max CUDA math operations on float/double types to
propagate NaNs: if either operand is NaN, the result should be NaN.
TODO: float16/bfloat16 need similar change.
### Motivation
Currently, results differ between the CPU and CUDA implementations of
the floating point Min and Max operators: the CPU operators correctly
return NaN results if either operand is NaN. This PR updates the CUDA
implementations to conform with this correct behavior.
See the the issue and comments raised
[here](https://github.com/onnx/onnx/issues/6003).
### Context
Same behavior in numpy, torch and Java:
```
>>> numpy.min([numpy.NAN, 1])
nan
>>> numpy.max([numpy.NAN, 1])
nan
>>> torch.min(torch.tensor([1, float('nan')]))
tensor(nan)
>>> torch.max(torch.tensor([1, float('nan')]))
tensor(nan)
```
C languguage [fmin](https://en.cppreference.com/w/c/numeric/math/fmin)
and [fmax](https://en.cppreference.com/w/c/numeric/math/fmax) has
different behavior:
```
fmax(NaN,1) = 1
fmin(NaN,1) = 1
```
https://grouper.ieee.org/groups/msc/ANSI_IEEE-Std-754-2019/background/minNum_maxNum_Removal_Demotion_v3.pdf
https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2273.pdf
