🚀 Feature

Currently, we have scatter_ and scatter_add_.

However, in use-cases such as Graph Neural Networks, it's common to wanting to do other reductions than add.
As a case in point, @rusty1s has written https://github.com/rusty1s/pytorch_scatter which has add, sub, mean, max, min, std, mul, div.

The kernels are there, and they are minor variations of the existing scatter kernel, see here.

We should provide this in PyTorch natively, ideally has a reduce='add' keyword argument in scatter_, or if one wants it to be more explicit, scatter_reduce_(..., mode='mean')

Motivation

Graph Neural Nets are becoming pretty common.

cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser @anjali411 @bhosmer @smessmer @ljk53 @ailzhang @nikitaved @pearu @cpuhrsch @IvanYashchuk