[Numpy][Gluon] Multiple gluon loss functions do not support numpy

[sxjscience]

For example, some losses won't support setting batch_axis, e.g.:
https://github.com/apache/incubator-mxnet/blob/4b86c32832a994e76b97dfc58c8a672db87e721d/python/mxnet/gluon/loss.py#L299-L302

Also, there are cases like HuberLoss that hasn’t supported numpy: https://github.com/apache/incubator-mxnet/blob/4b86c32832a994e76b97dfc58c8a672db87e721d/python/mxnet/gluon/loss.py#L604-L610

For the HuberLoss, it is tracked in #18603. However, there are more loss functions that do not support numpy.

@yzhliu @szha @leezu I mark this as 2.0 because we may keep the loss layers.

[szha]

I started working on this in #17196 but didn't get the time to finish yet. I will try to prioritize.

[szha]

The PR should be ready now.