Flux.destructure doesn't preserve RNN state

[avik-pal]

As pointed out here SciML/DiffEqFlux.jl#391 (comment), when reconstructing Recur, the state values are reverted to the initial state. A simple solution would be to do

mutable struct MyRecur{T}
  cell::T
  init
  state
end

function (m::MyRecur)(xs...)
  h, y = m.cell(m.state, xs...)
  m.state = h
  return y
end

Flux.@functor MyRecur

Flux.trainable(r::MyRecur) = Flux.trainable(r.cell)

EDIT: This doesn't work as expected

[DhairyaLGandhi]

Worth a PR, although it needs to be compatible with reset as well, so we don't move things off the GPU accidentally. In which case, are the arguments to functor doing anything?