ggml-webgpu: fast matrix-vector multiplication for i-quants

[SharmaRithik]

Overview

Adds fast WebGPU mat-vec implementations for all nine i-quant types (IQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ3_S, IQ4_NL, IQ4_XS). The kernels are added to mul_mat_vec.wgsl and selected through the existing use_fast dispatcher in ggml_webgpu_mul_mat.

Additional information

Numbers below are from test-backend-ops perf, comparing this branch vs. current master for the variant

MUL_MAT(type_a=<TYPE>,type_b=f32,m=4096,n=1,k=14336,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1)

across the nine i-quant types.

Intel Arc B580 (Mesa 25.2.8, Dawn 4654ba883e):

NVIDIA RTX 5080 (Dawn 4654ba883e):

AMD Radeon RX 7900 XT (Mesa 25.2.8, Dawn 4654ba883e):

Apple M2 (Dawn 4654ba883e):

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: no

[reeselevine]

Looks good! In terms of future work on i-quants, for you or anyone else who is interested in collaborating:

• We should add support to the shared memory loading for the matrix multiplication shaders, at which point we could fully remove the legacy mat-mul path.
• One optimization that some of the other backends do is load the i-quant tables into shared memory collaboratively, so that every thread doesn't have to maintain the giant array locally, which almost certainly leads to register pressure and spilling.