vulkan: Implement split_k for coopmat2 flash attention.

[jeffbolznv]

This is stacked on #12559.

When using group query attention, we have one workgroup per KV batch and this can be very few workgroups (e.g. just 8 in some models). Enable split_k to spread the work across SMs. This helps a lot when the KV cache is large.

before:
  FLASH_ATTN_EXT(hs=64,nh=8,nr=4,kv=4096,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                 11924 runs -   100.01 us/run -  33.55 MFLOP/run - 335.52 GFLOPS
  FLASH_ATTN_EXT(hs=128,nh=8,nr=4,kv=4096,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                 7455 runs -   139.64 us/run -  67.11 MFLOP/run - 480.59 GFLOPS
  FLASH_ATTN_EXT(hs=64,nh=8,nr=4,kv=8192,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                  5964 runs -   190.01 us/run -  67.11 MFLOP/run - 353.19 GFLOPS
  FLASH_ATTN_EXT(hs=128,nh=8,nr=4,kv=8192,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                 3730 runs -   268.44 us/run - 134.22 MFLOP/run - 499.98 GFLOPS
  FLASH_ATTN_EXT(hs=64,nh=8,nr=4,kv=16384,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                 2984 runs -   371.01 us/run - 134.22 MFLOP/run - 361.76 GFLOPS
  FLASH_ATTN_EXT(hs=128,nh=8,nr=4,kv=16384,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                1865 runs -   566.61 us/run - 268.44 MFLOP/run - 473.76 GFLOPS


Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -m C:\models\meta-llama-3-8b-instruct.Q4_K_M.gguf -p 0 -n 4096,8192,16384 -fa 0,1 --repetitions 1
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ------------: | -------------------: |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  0 |        tg4096 |         72.29 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  0 |        tg8192 |         67.03 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  0 |       tg16384 |         57.92 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  1 |        tg4096 |         66.44 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  1 |        tg8192 |         57.09 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  1 |       tg16384 |         43.84 ± 0.00 |

after:
  FLASH_ATTN_EXT(hs=64,nh=8,nr=4,kv=4096,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                 44715 runs -    22.66 us/run -  33.55 MFLOP/run - 1.48 TFLOPS
  FLASH_ATTN_EXT(hs=128,nh=8,nr=4,kv=4096,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                32802 runs -    31.55 us/run -  67.11 MFLOP/run - 2.13 TFLOPS
  FLASH_ATTN_EXT(hs=64,nh=8,nr=4,kv=8192,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                 29820 runs -    34.41 us/run -  67.11 MFLOP/run - 1.95 TFLOPS
  FLASH_ATTN_EXT(hs=128,nh=8,nr=4,kv=8192,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                20888 runs -    49.37 us/run - 134.22 MFLOP/run - 2.72 TFLOPS
  FLASH_ATTN_EXT(hs=64,nh=8,nr=4,kv=16384,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                17904 runs -    57.42 us/run - 134.22 MFLOP/run - 2.34 TFLOPS
  FLASH_ATTN_EXT(hs=128,nh=8,nr=4,kv=16384,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                6714 runs -   152.57 us/run - 268.44 MFLOP/run - 1.76 TFLOPS


Z:\github\jeffbolznv\llama.cpp\build\bin\RelWithDebInfo>llama-bench.exe -m C:\models\meta-llama-3-8b-instruct.Q4_K_M.gguf -p 0 -n 4096,8192,16384 -fa 0,1 --repetitions 1
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA GeForce RTX 4070 (NVIDIA) | uma: 0 | fp16: 1 | warp size: 32 | shared memory: 49152 | matrix cores: NV_coopmat2
| model                          |       size |     params | backend    | ngl | fa |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -: | ------------: | -------------------: |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  0 |        tg4096 |         72.31 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  0 |        tg8192 |         67.31 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  0 |       tg16384 |         57.95 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  1 |        tg4096 |         74.95 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  1 |        tg8192 |         70.85 ± 0.00 |
| llama 8B Q4_K - Medium         |   4.58 GiB |     8.03 B | Vulkan     |  99 |  1 |       tg16384 |         64.25 ± 0.00 |

[0cc4m]

LGTM.

Can you resolve the conflict?

[jeffbolznv]

Thanks. Conflicts were minor, I'll merge after it passes CI.