MMQ implementation for IQ4_KS_R4 and IQ5_KS_R4

[ikawrakow]

These two can use the more efficient block-of-32 MMQ GEMM kernels, so having MMQ implementation for them makes sense.

Sweep bench for LlaMA-3-8B on RTX-4080

Main branch IQ4_KS_R4

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.347	5910.02	4.052	126.34
2048	512	2048	0.325	6301.19	4.172	122.74
2048	512	4096	0.350	5848.94	4.417	115.92
2048	512	6144	0.378	5421.23	4.641	110.33
2048	512	8192	0.405	5052.95	4.863	105.28
2048	512	10240	0.432	4742.63	5.116	100.08
2048	512	12288	0.459	4459.86	5.302	96.57
2048	512	14336	0.486	4212.60	5.562	92.05

PR IQ4_KS_R4

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.281	7277.25	3.943	129.86
2048	512	2048	0.307	6674.86	4.159	123.12
2048	512	4096	0.335	6119.79	4.419	115.86
2048	512	6144	0.360	5681.17	4.648	110.16
2048	512	8192	0.389	5263.35	4.865	105.23
2048	512	10240	0.416	4927.52	5.118	100.05
2048	512	12288	0.443	4620.54	5.302	96.57
2048	512	14336	0.473	4330.80	5.557	92.14

Main branch IQ5_KS_R4

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.338	6052.15	4.674	109.55
2048	512	2048	0.326	6272.90	4.892	104.66
2048	512	4096	0.353	5800.11	5.149	99.43
2048	512	6144	0.380	5387.80	5.379	95.18
2048	512	8192	0.406	5041.40	5.597	91.48
2048	512	10240	0.434	4720.43	5.854	87.47
2048	512	12288	0.460	4451.96	6.037	84.81
2048	512	14336	0.489	4188.39	6.289	81.41

PR IQ5_KS_R4

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
2048	512	0	0.288	7118.13	4.669	109.66
2048	512	2048	0.313	6538.56	4.890	104.71
2048	512	4096	0.339	6034.98	5.149	99.44
2048	512	6144	0.368	5570.61	5.389	95.01
2048	512	8192	0.394	5193.25	5.619	91.12
2048	512	10240	0.422	4848.53	5.862	87.35
2048	512	12288	0.449	4562.94	6.045	84.70
2048	512	14336	0.479	4271.15	6.297	81.30