sweeps: RTX 5080 16GB row, plus a PDL sync bug in pr-ptq1-mmv on Blackwell

[LamplighterPaul]

The 16GB row, as promised on X, on an RTX 5080: the first Blackwell card in sweeps/. It found a bug on the way.

Bug: pr-ptq1-mmv at 285542d serves garbage on sm_120. llama-bench looks fine (97 tok/s), but in llama-server the text breaks a few tokens in, into !!!! or ////, for text and vision alike. GGML_CUDA_DISABLE_GRAPHS=1 or GGML_CUDA_PDL=0 makes it correct again. mul_mat_vec_ptq1_0_pt goes through ggml_cuda_kernel_launch, which opts into PDL on Hopper and newer, but the kernel never calls ggml_cuda_pdl_sync(), so it can read vy before the q8_1 quantize kernel has finished writing it. One line fixes it, at no speed cost:

     float      * GGML_CUDA_RESTRICT dst = dst_ptr;
+    ggml_cuda_pdl_sync();
+
     extern __shared__ float partials_dyn[];

Ampere and Ada never take the PDL path, which is why the 3060, 3060 Ti and 4070 rows are clean. test-backend-ops passes every PTQ1_0 MUL_MAT case with the unfixed build, because it runs each op on its own. PrismML-Eng/llama.cpp#218 carries the same kernel, so it will hit every Hopper and Blackwell user once merged. I'm happy to open that as a PR on your fork if it helps.

Numbers (llama-bench r=3, q4_0 K/V, fa on, stock b10685 / newest release b10735 / kernel+fix):

test	b10685	b10735	kernel+fix
tg128	87.58	87.51	96.92 (+10.8%)
tg128 @ d131072	41.96	41.82	44.19 (+5.7%)
pp512	995	2256	2179

Decode gains less than on the 3060: at 960 GB/s the stock kernel is less bandwidth-starved.

VRAM: the full 262144 window fits with the vision tower: 12,684 MiB for the server, with 2.4 GB still free on a card that also drives the desktop. 16gb-vision.sh could move from 131072 to 262144; I left the script alone for you to decide.

Full ladder, verbatim llama-bench output, the isolation table and the build notes are in sweeps/rtx5080-16gb.md. Every number comes from runs on this machine.

[LamplighterPaul]

Fix opened against pr-ptq1-mmv: sudoingX/llama.cpp#1. Also flagged on PrismML-Eng/llama.cpp#218.