Support for RNGs in dropout

[ablaom]

Now that we have FluxML/Flux.jl#1849 (comment), there are some things we could do:

• The rng argument for the Short builder should be passed to the Dropout layer at

  MLJFlux.jl/src/builders.jl

  Line 61 in b5a9c41

  Flux.Dropout(builder.dropout),

• Use Short instead of the test-local builder Short2 in tests and remove Short2:

  MLJFlux.jl/test/regressor.jl

  Line 6 in b5a9c41

  # TODO: replace Short2 -> Short when

• 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