The numerically stable version of logsumexp is simple but often useful. Thoughts on including this in PyTorch?

My thought is that it'd follow a call signature(s) that's similar to that of sum, but logsumexp wouldn't have the out keyword, and it'd also use the NumPy-consistent keepdims keyword instead of sum's keepdim keyword. So

def logsumexp(inputs, dim=None, keepdims=False):
    ...