vulkan: fix SSM_CONV crash on multi-GPU (#20462)

[ProgenyAlpha]

Reverts the 2D workgroup tiling (32x16) from #20379 which causes vk::DeviceLostError on multi-GPU RADV setups. Keeps the vec4 dot product fast path for nc=4.

The 2D tiling reduced workgroup launch overhead at large ubatch sizes, but triggers a driver-level fault on multi-GPU configurations (confirmed on dual 7900 XTX, bisected to 40c550d).

Based on master at 983df14.

Fixes #20462

Test plan

• @itterative: confirm crash reproduces on master (983df14) with multi-GPU
• @itterative: confirm this branch fixes the crash
• Single-GPU SSM_CONV test-backend-ops pass (verified, 45/45)

[itterative]

Tested against the pr, and it seems to be failing (same as master). Master fails at the same point as the second log file (vulkan-issue-0a2088a-2.txt).

$ llama-cli --version
ggml_vulkan: Found 2 Vulkan devices:
ggml_vulkan: 0 = Radeon RX 7900 XTX (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: KHR_coopmat
ggml_vulkan: 1 = Radeon RX 7900 XTX (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: KHR_coopmat
version: 8325 (0a2088a8c)
built with GNU 13.3.0 for Linux x86_64

Ran it twice after setting the temperature to 0. Here are the logs:
vulkan-issue-0a2088a.txt
vulkan-issue-0a2088a-2.txt

Some other notes:

• I tried to run llama-bench with context size 4100 (basically, 4 trailing tokens) and neither this pr, nor master fail
• there's a performance regression for tg128 (master is 90.70 t/s and this pr is 76.19 t/s) and I've seen this when running llama-server as well (pp4100 was approximately the same)

@ProgenyAlpha is there a way I could give you better stack traces for the failing vulkan shaders?

[ProgenyAlpha]

Closing, the revert didn't fix it. @itterative if you want to keep debugging this, the Vulkan debug output might help narrow down which op is faulting. You can enable it by adding target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG) to ggml/src/ggml-vulkan/CMakeLists.txt and rebuilding.

[ProgenyAlpha]

Correction: easier way to enable it is just the cmake flag:

cmake -B build -DGGML_VULKAN=ON -DGGML_VULKAN_DEBUG=ON

No need to edit CMakeLists.txt.