Port gamma and dirichlet implementations to CUDA

[fritzo]

@gchanan suggests:

we have pointwise support in ATen now for CUDA. See https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/cuda/TensorCompare.cu as an example. I'd put it in the same directory in Distributions.cu.
[the build system] should pick up any .cu file there.
If you want half (math) precision support in CUDA, this PR pytorch#4322 does that for embedding, i.e. adds AccumulateTypes, defines the common operators on Half so you don't have to write THCNumerics templates everywhere.

[fritzo]

@rachtsingh has already started a CUDA implementation of the Gamma sampler.

[rachtsingh]

Will land either with or just after #58

[rachtsingh]

Discussing here: pytorch#4839