vulkan: optimize and reenable split_k

[jeffbolznv]

Use vector loads when possible in mul_mat_split_k_reduce. Use split_k when there aren't enough workgroups to fill the shaders.

Split out from #10206.

I did a quick touch test to verify split_k helps the non-coopmat shaders as well:

before:
  MUL_MAT(type_a=f32,type_b=f32,m=128,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                    426 runs -  2600.37 us/run - 469.76 MFLOP/run - 180.65 GFLOPS
  MUL_MAT(type_a=f32,type_b=f32,m=256,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                    428 runs -  2569.83 us/run - 939.52 MFLOP/run - 365.60 GFLOPS
  MUL_MAT(type_a=f32,type_b=f32,m=384,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                    426 runs -  2579.22 us/run -   1.41 GFLOP/run - 546.40 GFLOPS
  MUL_MAT(type_a=f32,type_b=f32,m=512,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                    432 runs -  2582.09 us/run -   1.88 GFLOP/run - 727.72 GFLOPS

after:
  MUL_MAT(type_a=f32,type_b=f32,m=128,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                   1704 runs -   664.08 us/run - 469.76 MFLOP/run - 707.39 GFLOPS
  MUL_MAT(type_a=f32,type_b=f32,m=256,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                   1605 runs -   656.67 us/run - 939.52 MFLOP/run -   1.43 TFLOPS
  MUL_MAT(type_a=f32,type_b=f32,m=384,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                   1562 runs -   659.93 us/run -   1.41 GFLOP/run -   2.14 TFLOPS
  MUL_MAT(type_a=f32,type_b=f32,m=512,n=128,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                   1512 runs -   678.08 us/run -   1.88 GFLOP/run -   2.77 TFLOPS

[0cc4m]

Thank you! Pretty good improvement even without coopmat. I should have retested it myself. But I wouldn't have thought of the vector + scalar load in one shader thing, at best I'd have created a separate vector version.