Kmin removes infinite values

[mglisse]

The Kmin reduction replaces infinite values with the largest finite float:

import torch
from pykeops.torch import LazyTensor
a=torch.tensor([1e20])
b=LazyTensor(a)
c=b*b
print(c.Kmin(1,dim=0))
print(c.sum_reduction(dim=0))

tensor([[3.4028e+38]]) tensor([[inf]])

We were trying to test for inf or nan to detect overflow, and this broke the detection.

I am using pykeops 1.5 with torch 1.10.0+cu113.