Fix derived CuTracedArray construction in ReactantCUDAExt

[Copilot]

ReactantCUDAExt still referenced CUDA-internal derived-array helpers that do not exist for CuTracedArray. As a result, reshaping traced CUDA arrays in compiled kernels could fail on an undefined _derived_array path.

• Derived CuTracedArray reconstruction

  • add a raw-pointer constructor for CuTracedArray
  • add a local _derived_array helper that rebuilds traced CUDA arrays from the existing pointer plus the new shape encoded in the type parameters

• Remove stale CUDA leftovers from reshape/reinterpret

  • route Base.reshape through the local derived-array helper
  • update Base.reinterpret to use the same reconstruction path instead of accessing CUDA CuDeviceArray fields that CuTracedArray does not have

• Regression coverage

  • add a CUDA integration test that reshapes a traced array inside an @cuda kernel and checks a value that depends on Julia’s column-major reshape semantics

function reshape_kernel!(out, x)
    xr = reshape(x, 2, 2)
    @inbounds out[1] = xr[1, 2]
    return nothing
end

• Fixes Derived array not found in CUDAExt #3204

[ptiede]

@copilot can we not just use the _derived_array from CUDA.jl directly?

[Copilot]

@copilot can we not just use the _derived_array from CUDA.jl directly?

Yes — I switched this to use CUDA._derived_array when that helper exists, with a guarded import so the extension still loads if the symbol is absent. Addressed in adc7e6b.