Improvements to batched_mul, including PermutedDimsArray

[mcabbott]

I think that we can allow batched_gemm! to be called on many (but not all) PermutedDimsArrays, and this is generally much faster than first calling permutedims. This PR is an attempt to implement this. It also extends batched_mul! to take α, β scales like mul!.

EARLIER:

• It adds methods to treat PermutedDimsArray{<:Number,3,(2,1,3)} as equivalent to BatchedTranspose (see batched_transpose causes a 'Need an adjoint for constructor NNlib.BatchedTranspose' error Zygote.jl#552) and to allow batched_adjoint ∘ batched_transpose to be trivial on real-valued arrays.
• It deletes the method for copy(::BatchedAdjoint) etc, from implementation for batch-wise matrix multiplication #100, to return an Array like copy(::Adjoint).

[mcabbott]

This has changed a lot. The current form uses https://github.com/JuliaMatrices/ArrayLayouts.jl to keep track dimensions & strides, via traits UnitStrideFirst and UnitStride{2}, added there in JuliaLinearAlgebra/ArrayLayouts.jl#12 (which not yet merged).

See also #191 for an alternative approach.

Either of these can be made to work with JuliaGPU/CuArrays.jl#664, which will similarly allow permutations.

[mcabbott]

This could possibly now be done with https://github.com/SciML/ArrayInterface.jl instead of https://github.com/JuliaMatrices/ArrayLayouts.jl . But I'm hesitant to add dependencies (which would probably need to be added to CUDA.jl too), and I think #191 is simpler lower-tech solution.