invert h and x arguments in RNNs

[CarloLucibello]

In order to cause less breakage and avoid confusion such as in
https://discourse.julialang.org/t/dimension-mismatch-in-flux-jl-rnn-example/122410
this PR reinstates the previous ordering for the arguments to a cell, i.e. cell(h, x) instead of the cell(x, h) we have on master.

[CarloLucibello]

I'm not sure we should do this, I like more the cell(x, h) convention, and also pytorch and lux do the same.
It is true that we have cell(h, x) on v0.14, but the cells are not documented and exported, they only casually appear in the recurrent layers' docstrings https://fluxml.ai/Flux.jl/stable/reference/models/layers/#Flux.RNN

[MartinuzziFrancesco]

Weighing in on this, I do think that the cell(x, h) convention is more intuitive and, as you pointed out, in line with other libraries. Also the rnn changes in 0.15 are breaking already, so all examples should be adapted anyway I assume