vulkan: Use mul_mat_vec_id for small values of n

[jeffbolznv]

Change ggml_vk_mul_mat_vec_id_q_f16 to loop over the batch dimension and update the indexing calculations in get_offsets.

Mat-vec is faster than mat-mat for small values of n. We don't get the same reuse of the weights as in the non-ID path, but with this the cost is linear in n rather than n>1 being far slower than n==1.

Perf on 5090:

test-backend-ops.exe perf -o MUL_MAT_ID -p mxfp4

before:

  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=1,k=2880):                75400 runs -    13.37 us/run -  66.36 MFLOP/run -   4.96 TFLOPS
  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=4,k=2880):                 5278 runs -   190.94 us/run - 265.42 MFLOP/run -   1.39 TFLOPS
  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=8,k=2880):                 4347 runs -   239.34 us/run - 530.84 MFLOP/run -   2.22 TFLOPS
  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=512,k=2880):               1878 runs -   533.28 us/run -  33.97 GFLOP/run -  63.71 TFLOPS

after:

  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=1,k=2880):                73892 runs -    13.57 us/run -  66.36 MFLOP/run -   4.89 TFLOPS
  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=4,k=2880):                21112 runs -    48.15 us/run - 265.42 MFLOP/run -   5.51 TFLOPS
  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=8,k=2880):                10395 runs -    96.23 us/run - 530.84 MFLOP/run -   5.52 TFLOPS
  MUL_MAT_ID(type_a=mxfp4,type_b=f32,n_mats=32,n_used=4,b=0,m=2880,n=512,k=2880):               1941 runs -   515.45 us/run -  33.97 GFLOP/run -  65.91 TFLOPS

This came up when testing #18892. There are small batches that dominate the runtime:

before:
MUL_MAT_ID mxfp4 m=2880 n=4 k=2880 n_expert=32 batch=4: 71436 x 218.869 us = 1.56352e+07 us (1212.48 GFLOPS/s)

after:
MUL_MAT_ID_ADD_ID MUL_MAT_ID mxfp4 m=2880 n=4 k=2880 n_expert=32 batch=4: 71508 x 53.676 us = 3.83833e+06 us (4943.93 GFLOPS/s)

That test also shows a lot of time in small batches of FA, but I'll fix that separately.