Add option to shuffle data after each epoch

[ablaom]

Probably shuffling without breaking batches is fine. I suggest adding an rng hyperparameter as well, which is interpreted as MersenneTwister(rng) if rng is an integer (as elsewhere in MLJ), and falls back to Random.GLOBAL_RNG This could ultimately be passed to the chain intitializers, although Flux does not currently make this easy (FluxML/Flux.jl#1335).

I also suggest the following handling for RNG here (and more generally) to help with reproducibility: The fit method creates a deep copy of the RNG, which then gets mutated as various rand(rng, ...) calls are made. The final state is then output to cache so that update can carry on with the mutated RNG in the case of a warm restart. In this way,

(i) multiple warm-restarts behave the same as training all in one go (modulo the chain initialisation problem), even if the original RNG gets used somewhere else in between restarts; and

(ii) By specifying a concrete RNG at model construction time, cold-restarts (with, eg, fit!(mach, force=true)) give the same behaviour every time.

@ayush-1506

[ayush-1506]

Will take this up. self-assigning.

[ablaom]

@AYUSH-150 We may want to deal with #160 first, as I'm needing this now.