[AzamatB]

unsafe_free! of a GPU array doesn't free its memory if the array was reduced first. The whole-array kernels read their source through @view src[begin:end]. For a CuArray, GPUArrays builds that view with derive, which makes a new CuArray holding a reference to the source's memory until the finalizer runs. Since GPUArrays 12 sends Base reductions through AK, this affects plain sum, maximum, count and friends, and it applies to commutative=false and dims reductions into a single output too:

a = CuArray(rand(Float32, 10^6))
sum(a)
CUDA.synchronize(); used = CUDA.used_memory()
CUDA.unsafe_free!(a); CUDA.synchronize()
used - CUDA.used_memory()   # 0 on main, 4000000 with this PR

The kernels index their source linearly from 1, so for a source that already indexes that way (IndexLinear, no offset axes), the view was just another name for the same elements. Such sources are now passed through unchanged. Sources with Cartesian indexing still get the view: it reshapes them, so linear indices are decoded with multiplicative inverses. A reduction along dims into a single output of a wrapper that covers a whole dense array, e.g. sum(transpose(v); dims=2), viewed the wrapped array the same way, and now passes that array instead. Leaving the view in and having CUDA.jl free it early isn't an option, because a derived array is supposed to keep its parent alive. The fix is for AK to not create the view in the first place.

Cost: an N-d array now compiles its own kernel specialization instead of sharing the vector's. On an RTX 5080, the first mapreduce(f, +, ::CuMatrix) after the same call on a CuVector takes ~250 ms instead of ~100 ms. The generated code is unchanged apart from computing the length from both dimensions, and sum/maximum of a CuVector and a CuMatrix (10^4 and 10^8 elements) run at the same speed as before.

Other operations (reductions along dims, scans, sorts, findall, broadcast sources) don't view the user's array and were already fine. The new CUDA test checks used_memory() after unsafe_free! for a vector, a matrix with dims=(1, 2), commutative=false and the transpose case, and fails on main. A wrapper that covers only part of its parent still gets a view, which would take an offset in the kernels to avoid.