vulkan: optimize coopmat2 iq2/iq3 callbacks

[jeffbolznv]

I've only looked at the directed perf tests, and there may still be more to do. But this gets to the same order of magnitude as Q4_K (included for comparison):

before:
  MUL_MAT(type_a=q4_K,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                  626 runs -  1597.97 us/run -  60.13 GFLOP/run -  37.63 TFLOPS
  MUL_MAT(type_a=iq2_xxs,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):               276 runs -  3629.79 us/run -  60.13 GFLOP/run -  16.57 TFLOPS
  MUL_MAT(type_a=iq2_xs,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                252 runs -  3981.24 us/run -  60.13 GFLOP/run -  15.10 TFLOPS
  MUL_MAT(type_a=iq2_s,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                 338 runs -  2970.81 us/run -  60.13 GFLOP/run -  20.24 TFLOPS
  MUL_MAT(type_a=iq3_xxs,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):               284 runs -  3525.55 us/run -  60.13 GFLOP/run -  17.06 TFLOPS
  MUL_MAT(type_a=iq3_s,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                 330 runs -  3039.16 us/run -  60.13 GFLOP/run -  19.78 TFLOPS

after:
  MUL_MAT(type_a=q4_K,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                  626 runs -  1598.70 us/run -  60.13 GFLOP/run -  37.61 TFLOPS
  MUL_MAT(type_a=iq2_xxs,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):               632 runs -  1585.54 us/run -  60.13 GFLOP/run -  37.92 TFLOPS
  MUL_MAT(type_a=iq2_xs,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                552 runs -  1811.94 us/run -  60.13 GFLOP/run -  33.19 TFLOPS
  MUL_MAT(type_a=iq2_s,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                 496 runs -  2016.71 us/run -  60.13 GFLOP/run -  29.82 TFLOPS
  MUL_MAT(type_a=iq3_xxs,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):               670 runs -  1494.11 us/run -  60.13 GFLOP/run -  40.24 TFLOPS
  MUL_MAT(type_a=iq3_s,type_b=f32,m=4096,n=512,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3]):                 670 runs -  1496.61 us/run -  60.13 GFLOP/run -  40.18 TFLOPS

[0cc4m]

LGTM