Shall we add half_t support to Random<cpu, DType>

[KiddoZhu]

Member functions in Random<cpu, DType> are not overload for half_t type. As a result, it fails when we use Random<cpu, half_t> to generate a tensor. This is quite inconvenient for generating noise inside a mxnet operator that is templated by DType.

Shall we add half_t support to Random<cpu, DType>?

[BiranLi]

I think Random<gpu, DType> is important too.
It's not supported by curand function in CUDA9. But I support U.

[szha]

This code base has been donated to the Apache MXNet project per #373, and repo is deprecated. Future development and issue tracking should continue in Apache MXNet.