[maleadt]

The scans have the same problem #204 fixes for the reductions (JuliaGPU/CUDA.jl#3516). Each scan kernel keeps a tile of block_size * items_per_thread partial results in local memory, plus one partial result per thread, and the default items_per_thread never goes below 1. With the default 256 threads, any element over 64 bytes overruns the 32 KiB budget, and any element over 96 bytes fails to compile on CUDA devices before Blackwell (48 KiB of static shared memory). That covers cumsum of ForwardDiff duals with 16 partials (136 bytes) and wider structs.

• Large element types: an unset block size is now lowered to the largest power of two whose tile of one item per thread, plus the per-thread totals, fits local_mem_bytes. As in Fit the reductions' partial results in local memory for large element types #204 (CUB's MemBoundScaling), the size is taken from the partial-result type, which is an Optional for operators without a neutral element. The block size never goes below 2, since ScanPrefixes needs tiles of at least two elements. An explicit block_size is kept as given. This covers ScanPrefixes, SliceScan and everything built on them (accumulate!, cumsum, cumprod). The :rows slices already fit the budget, and findall's block scans work on Int32 counts.
• One value per sub-group: when a block scan uses sub-groups, it now keeps one slot per sub-group in local memory instead of one per thread, the way Combine the values of a block with sub-group reductions #205 does for the reductions. Each thread passes its value to block_exclusive_scan directly. This removes two barriers, and for large elements it shrinks the scan kernels' shared memory further (on sm_75, the largest scan kernel went from 32 KiB to 20 KiB). The tuning still sizes tiles as if there were one value per thread, because backends without sub-groups, and CPUs, need that. The host backend's tests turn sub-groups on, so this path is also checked under PoCL.

On sm_75, before this change, 38 of the 83 scan kernels compiled for structs of 64 to 2080 bytes and 16-partial duals (whole arrays and along dims, with and without a neutral element) were rejected by ptxas -arch=sm_75. After it, all 117 assemble.

Kernel time on an RTX 5080, in µs. Each value is the device time of one call, best of 20; the table shows the median of 3 interleaved runs per side, with the range in parentheses:

case	before	after
cumsum F32 1e6	7.6 (7.6–7.6)	7.6 (7.6–7.7)	-0.8%
cumsum F32 1e8	1387.9 (1385.2–1388.9)	1388.2 (1386.4–1388.4)	+0.0%
cumsum F64 1e7	265.2 (264.8–266.6)	263.7 (263.4–264.5)	-0.5%
cumsum F32 1e7, no neutral	92.2 (91.2–92.3)	88.7 (87.9–89.2)	-3.8%
cumsum F32 4096² dims=1	140.1 (139.7–140.2)	141.4 (140.6–141.6)	+1.0%
cumsum F32 4096² dims=2	238.7 (238.5–239.4)	239.9 (238.6–240.9)	+0.5%
cumsum F32 2×1e7 dims=2	274.1 (273.3–274.3)	272.3 (271.9–272.4)	-0.7%
cumsum F32 1e6×8 dims=1	43.3 (43.2–43.3)	41.5 (41.5–42.2)	-4.0%
cumsum F64 1000×1e4 dims=1	175.6 (175.3–176.0)	176.2 (176.0–176.9)	+0.3%
findall F32 1e7	147.8 (147.4–147.9)	147.8 (147.2–148.2)	-0.0%
scan 32-byte struct 1e6	54.7 (54.6–55.0)	51.9 (51.6–52.0)	-5.0%
scan 96-byte struct 1e6	329.3 (328.5–329.5)	325.8 (325.7–326.4)	-1.0%
scan 96-byte struct 3×1e5 dims=2	124.9 (124.7–125.5)	111.7 (110.5–112.0)	-10.5%
cumsum Dual16 1e6	1716.7 (1690.3–1720.0)	1214.7 (1214.3–1214.9)	-29.2%
cumsum Dual16 3×1e5 dims=2	603.4 (600.3–604.1)	472.5 (472.5–472.8)	-21.7%

(The large types do compile on Blackwell before this change, because it allows about 99 KiB of shared memory.)