Some cat improvements

[mcabbott]

Before:

julia> let 
       x,y,z = [onehotbatch(rand(1:10, 100), 1:10) for _ in 1:3]
       @btime vcat($x, $y, $z)
       @btime hcat($x, $y, $z)
       end;
  4.214 μs (9 allocations: 6.41 KiB)
  120.640 ns (4 allocations: 2.16 KiB)

julia> reduce(hcat, [onehot(2, 1:3), onehot(1, 1:3)])
3×2 Matrix{Bool}:
 0  1
 1  0
 0  0

julia> cat(onehot(2, 1:3), onehot(1, 1:3); dims=(1,2))
ERROR: TypeError: in keyword argument dims, expected Int64, got a value of type Tuple{Int64, Int64}

After:

julia> let 
       x,y,z = [onehotbatch(rand(1:10, 100), 1:10) for _ in 1:3]
       @btime vcat($x, $y, $z)  # now just calls invoke, instead of collecting the pieces first
       @btime hcat($x, $y, $z)  # better code, avoids a bad reduce(hcat, ::Tuple)
       end;
  3.406 μs (3 allocations: 3.08 KiB)
  61.650 ns (2 allocations: 1.25 KiB)

julia> reduce(hcat, [onehot(2, 1:3), onehot(1, 1:3)])  # calls stack
3×2 OneHotMatrix(::Vector{UInt32}) with eltype Bool:
 ⋅  1
 1  ⋅
 ⋅  ⋅

julia> cat(onehot(2, 1:3), onehot(1, 1:3); dims=(1,2))
6×2 Matrix{Bool}:

Closes #32, closes FluxML/Flux.jl#1596