[Numpy] Whether to support 'exclude' flag in numpy operators

[sxjscience]

In the old ndarray API, we support mx.nd.mean(x, axis=1, exclude=True), in which the meaning is to sum up the axes except axis=1. We do not have the exclude option in mx.np.mean() so it will be good to have.

See the following code snippet for an example:

https://github.com/apache/incubator-mxnet/blob/74430a92f5232808a169c9e238a419789a5c2fda/python/mxnet/gluon/loss.py#L149-L152

Another option is to directly switch to use DeferedCompute so there is no need to add this option.

[szha]

While this design made a difference when we needed to program in the symbolic way, since in 2.0 the array interface provides eager type inference, our array can behave just like numpy ndarray so this design shouldn't be required anymore. As such, I'm leaning towards that it's still better to be as consistent with the numpy API as possible in the np module.