permutedims function not supported for Dagger.DArray

[jofrevalles]

Hello,
During my usage of DArray in the Dagger.jl library, I encountered an inconsistency between how transpose and permutedims functions operate. While the transpose function works perfectly with DArray, it appears that permutedims is not currently supported.

Here's a minimal example that demonstrates the issue:

julia> darray = distribute(rand(4, 4), Blocks(2, 2))
DArray{Any, 2, typeof(cat)}(4, 4)

julia> transpose(darray) # works fine
Dagger.Transpose{Any, 2}(4, 4)

julia> transpose(darray) |> compute |> collect
4×4 Matrix{Float64}:
 0.167282  0.111796   0.139091   0.76079
 0.970018  0.962297   0.0806159  0.750641
 0.44158   0.213501   0.929178   0.0804493
 0.980127  0.0202904  0.873157   0.249399

julia> permutedims(darray, (2, 1))
4×4 Matrix{Any}:
 Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)
 Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)
 Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)
 Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)  Dagger.GetIndex{Any, 2}(1, 0, 1, 0)

The result of permutedims(darray, (2, 1)) is a matrix of Dagger.GetIndex objects, and attempting to compute the result leads to an error, as shown here:

julia> permutedims(darray, (2, 1)) |> compute
ERROR: MethodError: no method matching compute(::Context, ::Matrix{Any}; options=nothing)
Closest candidates are:
  compute(::Context, ::DArray; persist, options) at ~/.julia/packages/Dagger/vNUsP/src/array/darray.jl:231
  compute(::Context, ::Thunk; options) at ~/.julia/packages/Dagger/vNUsP/src/compute.jl:27
  compute(::Any; options) at ~/.julia/packages/Dagger/vNUsP/src/compute.jl:5
  ...

I couldn't find any documentation addressing this issue, so I thought it might be worth bringing up as a potential enhancement for DArray. The ability to perform arbitrary dimension permutations would enhance the usability of DArray, particularly for tensor computations.

Thanks for your time and all your work on this library!

Best,
Jofre