[maleadt]

Every reduction kernel keeps a partial result per thread in local memory, in blocks of 256 threads by default, without checking the element size. Elements larger than 128 bytes exceed the 32 KiB budget that the sorts and scans keep to, and those larger than 192 bytes fail to compile on CUDA devices before Blackwell (48 KiB of static shared memory; about 396 bytes on Blackwell). ForwardDiff duals with 16 partials, or nested ones, are that large: JuliaGPU/CUDA.jl#3516.

As CUB's MemBoundScaling, an unset block size is now lowered to the largest power of two whose partial results fit the tuning's new local_mem_bytes. The size is that of the partial-result type, an Optional for operators without a neutral element, which resolve_reduce takes as an optional argument like resolve_scan. The second pass of split reductions stays within the first pass's block size. An explicit block_size is kept as given.

Tests: resolution of the block size for elements of 4 bytes to 64 KiB, and reductions of 648- and 2080-byte elements (blocks of 32 and 8) along every launch shape, which fail to compile without this change. On AMDGPU, ordered reductions of the larger size need AMDGPU_LLVM_Backend_jll 23.1.3+1 (JuliaPackaging/Yggdrasil#15464, backporting llvm/llvm-project#239163): LLVM's AMDGPU back-end crashed on the kernel's divergent endpgm once it outgrew the branch range.

Sub-group reductions, which shrink the kernels' local memory further, follow in a stacked PR. Ordered reductions of elements of more than 32 words also need JuliaGPU/KernelAbstractions.jl#947, so this should be merged after KernelInterface 0.4.5 is registered.