CUDA: faster IQ3_K, IQ3_KS, IQ3_K_R4

[ikawrakow]

This PR is a follow up of #713, and applies a similar trick to 3-bit quants that need a table lookup (IQ3_K, IQ3_KS, IQ3_K_R4).

Here some performance comparisons to the main branch for LlaMA-3.1-8B-Instruct on RTX-4080

model	test	t/s (main)	t/s (PR)	Speedup
llama 8B IQ3_KS	pp512	8096.66 ± 36.79	8507.01 ± 44.55	1.051
llama 8B IQ3_K	pp512	6705.65 ± 29.94	7027.30 ± 36.92	1.048
llama 8B IQ3_K_R4	pp512	6503.14 ± 46.09	7062.74 ± 38.44	1.086
llama 8B IQ3_KS	tg128	148.14 ± 0.32	154.84 ± 0.08	1.045
llama 8B IQ3_K	tg128	144.16 ± 0.15	148.57 ± 0.15	1.031
llama 8B IQ3_K_R4	tg128	138.70 ± 0.02	145.49 ± 0.10	1.049