[Feature Request] Support CSRNDArray in kv_store

[zerolxf]

Description

In distributed training, communication takes up a lot of time. In order to compress the communication time, an optimization method is to make the gradient sparse.

But for now kv_store doesn't support push CSRNDArray values.
Although support for this feature may cause other problems, right?

For example, for CSRNDArray format

The format is designed for 2D matrices with a large number of columns, and each row is sparse

But kv_store probably transfers multidimensional arrays

[vdantu]

@mxnet-label-bot add [KVStore, Feature request, Question]

[eric-haibin-lin]

Yes! I am also thinking about doing this. I am curious - what's your use case?

[zerolxf]

I want to implement the optimizer described in the paper "Sparse Communication for Distributed Gradient Descent" , available at https://arxiv.org/abs/1704.05021

I tried to modify the value of NDArray directly through index, but it was a bit slow.