Support tuple dims in mapreduce_nd using CartesianIndices

[shreyas-omkar]

Support tuple dims in mapreduce_nd

Extends mapreduce/reduce to accept dims as a Tuple{Vararg{Int}} in addition to a single Int, matching the semantics of Base.mapreduce.

What changed

• mapreduce_nd signature updated to dims::Union{Int, Tuple{Vararg{Int}}}
• Replaced stride-arithmetic-based kernels with a CartesianIndices-based approach, splitting iteration into Rother (kept dims) and Rreduce (reduced dims), mirroring the approach in GPUArrays.jl's vc/ka_mapreducedim branch
• Single kernel pass, no intermediate allocations, works for any combination of dims

Testing

• All existing tests pass
• New tests added for dims=(1,2), (1,3), (2,3), (1,2,3) on 3D arrays
• Verified on CPU, Intel GPU (oneAPI/Iris XE), NVIDIA GPU (CUDA/RTX 3060)

[shreyas-omkar]

@maleadt and @christiangnrd, Please take a look.