Use AbstractRNG type

[wildart]

Use an Base.AbstractRNG as a super type for your RNG class and implement necessary methods to work with AbstractRNG derivatives. This would provide compatibility with Julia RNG interface.

[dfdx]

I didn't know Julia has AbstractRNG type, thanks. However, I don't think it's a good idea to bind CURAND functions to it. Here's a couple of reasons:

• original cuRAND doesn't support most operations that AbstractRNG should support (e.g. randexp, randperm. shuffle, sprand, etc.); probably we can add them using custom kernels, but it's not trivial at all
• most AbstractRNG-related methods operate on AbstractArray, and CudaArray doesn't implement it (for a good reason);
• two APIs are not consistent anyway, e.g. cuRAND's randn() takes mean and stddev parameters, while Julia's doesn't; cuRAND has curand_logn and curand_poisson, Julia instead has randexp, etc.

On other hand, it's worth to make API more "similar" to that of Julia, so I'll work it out today.