CUDA code refactoring

[CarloLucibello]

This moves all cuda related code inside the (already existing) CUDAint submodule.
Now the Flux module doesn't have a direct using CUDA statement.

Also

• reintroduces the has_cudnn check per comment cuda 4.0 compat #2177 (comment) (cc @maleadt)
• removes the v0.12 deprecations
• factorized the old train! compatibility methods in their own file deprecations_train.jl in order to clean up deprecations.jl

I'm sorry it predates a bit #2132, but maybe better to start this refactoring now while waiting for julia v1.9 to be released.

[codecov-commenter]

Codecov Report

Base: 86.02% // Head: 86.57% // Increases project coverage by +0.54% 🎉

Coverage data is based on head (00d95a9) compared to base (c5a691a).
Patch coverage: 79.24% of modified lines in pull request are covered.

📣 This organization is not using Codecov’s GitHub App Integration. We recommend you install it so Codecov can continue to function properly for your repositories. Learn more

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2182      +/-   ##
==========================================
+ Coverage   86.02%   86.57%   +0.54%     
==========================================
  Files          19       22       +3     
  Lines        1460     1452       -8     
==========================================
+ Hits         1256     1257       +1     
+ Misses        204      195       -9     

Impacted Files	Coverage Δ
src/Flux.jl	0.00% <ø> (ø)
src/cuda/utils.jl	0.00% <0.00%> (ø)
src/deprecations.jl	19.44% <ø> (-18.24%)	⬇️
src/functor.jl	97.22% <ø> (+6.74%)	⬆️
src/utils.jl	96.68% <ø> (+0.63%)	⬆️
src/cuda/functor.jl	79.31% <79.31%> (ø)
src/deprecations_train.jl	82.60% <82.60%> (ø)

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[CarloLucibello]

Merge?

[mcabbott]

What's gained by moving things around, now instead of once we figure out extensions (#2132)?

has_cudnn could be separate.

Removing deprecations should ideally happen right before 0.14.

The train deprecations would probably also go then. What's gained by moving them around (losing blame) and why along with CUDA?