Adding return state option to recurrent layers

[MartinuzziFrancesco]

Addressing the return_state as discussed in #2514.

PR Checklist

• Tests are added
• Entry in NEWS.md
• Documentation, if applicable

[CarloLucibello]

tests are failing

[MartinuzziFrancesco]

is this something that has to be addressed at the Functor level? the fail is at https://github.com/FluxML/Flux.jl/blob/master/test/test_utils.jl#L66 with

MethodError: no method matching RNN(::RNNCell{typeof(tanh), Matrix{Float64}, Matrix{Float64}, Vector{Float64}})

so I guess it's trying to rebuild the struct but it can't since it doesn't have the return_state parametrization information. I can add a workaround like

RNN(rnn::RNNCell) = RNN{false, RNNCell}(rnn)

but this wouldn't solve it at the level that it's needed. The tests would still fail for return_state = true

[CarloLucibello]

yes this needs a custom functor implementation. Maybe something like

Functors.functor(rnn::RNN{S}) = (; rnn.cell), ch -> RNN{S}(ch.cell)

(not tested)

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 65.53%. Comparing base (7a3eafd) to head (710fb10).
Report is 7 commits behind head on master.

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #2557       +/-   ##
===========================================
+ Coverage   32.50%   65.53%   +33.02%     
===========================================
  Files          34       34               
  Lines        2003     2083       +80     
===========================================
+ Hits          651     1365      +714     
+ Misses       1352      718      -634     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.