feat: use loop vectorization for faster cpu broadcasting

[avik-pal]

Updates

• uses LoopVectorization if possible for broadcasting on CPU.
• refactors fast_activation!! to call fast_broadcast!!. the former is deprecated.
• chainrules bypasses the LV impl completely.
  • For common activation functions, we see a >2x performance boost. Fallback performance still remains constant which is a good thing.
• faster broadcasting in general:
  • dropout functions

Current Issues

• CUDA CI time seems to have gone up
• AMDGPU fused_dense_bias_activation gradients fails for certain inputs

TODOs

• for enzyme we need a workaround with mixed forward mode.
  • bypasses Atomic fadd of vector EnzymeAD/Enzyme.jl#1635 doesn't allow LV.
  • hopefully we can temporarily work around Slow Broadcasting (compared to Zygote) EnzymeAD/Enzyme.jl#1434
• need some proper benchmarks showcasing the improvements. we don't need any upstream Lux changes so benchmarking here is quite easy.
• needs some testing for gradient correctness.
• special case for broadcasting bias for convolutions. right now falls back to FB
• faster broadcasting for other impls:
  • normalization functions
  • affine functionss
• do we want to add a public bias_activation API?

[avik-pal]

Branches have diverged too much. I will open a fresh PR for this.