hotfix LSTM ouput

[CarloLucibello]

The LSTM layer in Flux v0.15.0 returns a tuple of (h, c) where size(h)==(in_dim, timesteps, batch_size). This is in contrast with any other framework (pytorch, flax, lux), where only h is returned as an output. Due to this, we cannot stack LSTM in a Chain.

With this PR, the LSTM returns only h. While this change is breaking, v0.15 has only been around for a few days so we can consider the current behavior a bug introduced while doing the redesign in #2500, fix it now and tag a v0.15.3 as soon as possible.

cc @MartinuzziFrancesco

[codecov]

Codecov Report

Attention: Patch coverage is 80.00000% with 4 lines in your changes missing coverage. Please review.

Project coverage is 32.20%. Comparing base (428be48) to head (f96bd58).
Report is 2 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layers/recurrent.jl	80.00%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2547      +/-   ##
==========================================
- Coverage   32.44%   32.20%   -0.25%     
==========================================
  Files          34       34              
  Lines        1991     1972      -19     
==========================================
- Hits          646      635      -11     
+ Misses       1345     1337       -8     

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

[MartinuzziFrancesco]

I think the change makes sense in light of the other implementations. Pushing it as an quick hotfix shouldn't be an issue, as you said 0.15 is pretty new. This will also probably be superseded by closing #2514 when all the changes will be implemented, specifically adding num_layers to bring the implementation in line with Pytorch, Flax etc