remove DepthwiseConv type in favor of Conv

[CarloLucibello]

Implement @mcabbott's comment

Also adds the assertion

  @assert cin % groups == 0 "Input channel dimension must be divisible by groups."
  @assert cout % groups == 0 "Output channel dimension must be divisible by groups."

for standard convolutions.

Close #1667, close #459

[codecov-commenter]

Codecov Report

Merging #1921 (cab5f26) into master (57beb23) will increase coverage by 0.22%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1921      +/-   ##
==========================================
+ Coverage   86.43%   86.66%   +0.22%     
==========================================
  Files          18       18              
  Lines        1430     1417      -13     
==========================================
- Hits         1236     1228       -8     
+ Misses        194      189       -5     

Impacted Files	Coverage Δ
src/layers/show.jl	72.72% <ø> (ø)
src/outputsize.jl	84.21% <ø> (ø)
src/utils.jl	95.30% <ø> (ø)
src/layers/conv.jl	81.76% <100.00%> (+1.43%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 57beb23...cab5f26. Read the comment docs.

[CarloLucibello]

ready for review/merge

[mcabbott]

Should what's left move to deprecations.jl too?

[CarloLucibello]

I think it is good to have a convenience constructor like that, helps you understand what a depthwise convolution is

[ToucheSir]

Would you mind running some microbenchmarks on CPU pre- and post-PR? If there's basically no difference between the two, let's merge asap :)

[mcabbott]

Xref FluxML/NNlib.jl#395 perhaps. The present code has 3 layers of multi-threading, not sure how carefully it was ever benchmarked.

[CarloLucibello]

Performance here seems slightly better. We should remove the specialized depthwiseconv from NNlib.

This PR

julia> using Flux, BenchmarkTools

julia> x = randn(Float32, 128, 128, 32, 32);

julia> dconv = DepthwiseConv((3,3), 32 => 64)
Conv((3, 3), 32 => 64, groups=32)  # 640 parameters

julia> Threads.nthreads()
1

julia> @btime $dconv($x);
  169.259 ms (320 allocations: 265.55 MiB)

julia> Threads.nthreads()
4

julia> @btime $dconv($x);
  96.917 ms (366 allocations: 317.88 MiB)

On Master

julia> using Flux, BenchmarkTools

julia> x = randn(Float32, 128, 128, 32, 32);

julia> dconv = DepthwiseConv((3,3), 32 => 64)
DepthwiseConv((3, 3), 32 => 64)  # 640 parameters

julia> Threads.nthreads()
1

julia> @btime $dconv($x);
  180.530 ms (22 allocations: 265.51 MiB)

julia> Threads.nthreads()
4

julia> @btime $dconv($x);
  101.642 ms (37 allocations: 317.83 MiB)

[ToucheSir]

Great, we can move the perf discussion (mainly figuring out where that extra memory usage comes from) to NNlib.