feat: correct handling of wrapped arrays functionalities

[avik-pal]

fixes #339

The main trick here is to preserve ReshapedArray type. Applies the fix for the following types for the time being:

1. ReshapedArray
2. Adjoint
3. Transpose
4. Diagonal
5. PermutedDimsArray

Adds the following LinearAlgebra functionality

1. diag
2. diagm

Updated NNlibExt to handle wrappers correctly

[mofeing]

Maybe this can already be replaced by Ops.maximum?

[avik-pal]

The first argument here is being used directly with block arguments. It would be an unnecessary in-direction to go down the Ops route here

[mofeing]

it fixes my code!

using Tenet#master with this

using Tenet, Reactant, Adapt
julia> a = Tensor([1 0; 1 0], [:j, :i])
2×2 Tensor{Int64, 2, Matrix{Int64}}:
 1  0
 1  0

julia> b = Tensor([0 0; 1 1], [:i, :j])
2×2 Tensor{Int64, 2, Matrix{Int64}}:
 0  0
 1  1

julia> a + b # this is correct
2×2 Tensor{Int64, 2, Matrix{Int64}}:
 1  1
 1  1

julia> parent(a) + parent(b) # this is wrong
2×2 Matrix{Int64}:
 1  0
 2  1

before this PR

julia> ar = adapt(ConcreteRArray, a)
2×2 Tensor{Int64, 2, ConcreteRArray{Int64, 2}}:
 1  0
 1  0

julia> br = adapt(ConcreteRArray, b)
2×2 Tensor{Int64, 2, ConcreteRArray{Int64, 2}}:
 0  0
 1  1

julia> @jit ar + br # it's taking the parent array, not transposed
2×2 Tensor{Int64, 2, ConcreteRArray{Int64, 2}}:
 1  0
 2  1

after this PR

julia> @jit ar + br
2×2 Tensor{Int64, 2, ConcreteRArray{Int64, 2}}:
 1  1
 1  1

[mofeing]

everythings fine but would you mind adding some more tests for Ops.reshape? we probably missed the bug you're talking about

[mofeing]

ahhh i didn't think about using dynamic_gather in this way. i feared about using dynamic_slice for each diagonal element. nice!

[avik-pal]

The only additional CI failure will be resolved by #362