Support for RNGs in dropout

[ablaom]

Now that we have https://github.com/FluxML/Flux.jl/pull/1849#event-5967026617, there are some things we could do:

• [ ] The rng argument for the Short builder should be passed to the Dropout layer at https://github.com/FluxML/MLJFlux.jl/blob/b5a9c419bb3af4aae18fdbd3fb69b03497b6ea9e/src/builders.jl#L61
• [ ] Use Short instead of the test-local builder Short2 in tests and remove Short2: https://github.com/FluxML/MLJFlux.jl/blob/b5a9c419bb3af4aae18fdbd3fb69b03497b6ea9e/test/regressor.jl#L6
• [ ] Update this section of the README.md: https://github.com/FluxML/MLJFlux.jl#random-number-generators-and-reproducibility. We should add sentence explaining limitations of Dropout RNG's for GPU training (and copy @darsnack in the PR to check we have this correct)
• [ ] In the README.md example at https://github.com/FluxML/MLJFlux.jl#random-number-generators-and-reproducibility, include a Dropout layer with rng