Fp16 support for softmax

[eric-haibin-lin]

Currently, given fp16 inputs, nd.softmax/sym.softmax perform reduction in fp16, which losses precision. The reduction should be done in fp32 instead.

https://github.com/apache/incubator-mxnet/blob/32c9ca74839ae4d275bcf9a027ea0a711373be81/src/operator/nn/softmax-inl.h#L164-L202
pytorch reference:
https://github.com/zdevito/ATen/blob/a6cc4156fe4abc9e31f62f2bba1a2f68c58b77b7/aten/src/ATen/native/cuda/SoftMax.cu#L43-L55

[mxnet-label-bot]

Hey, this is the MXNet Label Bot.
Thank you for submitting the issue! I will try and suggest some labels so that the appropriate MXNet community members can help resolve it.
Here are my recommended labels: Feature

[eric-haibin-lin]

fixed in #14098