IQ2_XXS: much faster CPU prompt processing

[ikawrakow]

While experimenting with the trellis quants in PRs #505 and #511, I realized that CPU matrix multiplications (GEMM) for quants that are slow to unpack and make ready for int8_t dot products (as the trellis quants are) are much faster if one unpacks a given number of rows to, e.g., Q8_0_R8, and then uses the Q8_0_R8 x Q8_2_X4 GEMM to perform the multiplication with all columns of the right matrix.

This PR applies the approach of #505/#511 to IQ2_XXS (AVX2/Zen4 only). We get nearly 3X improvement in PP performance compared to IQ2_XXS on the main branch, and 2X compared to IQ2_XXS_R4!

The same approach can be used out-of-the-box for IQ3_XXS (left for a follow up PR).

IQ2_XS, IQ2_S and IQ3_S use blocks of 16, so one would need a new row-interleaved 8-bit type with blocks of 16 for those.

[ikawrakow]

Here some sweep-bench tables on a Ryzen-7950X CPU for LlaMA-3.1-8B

IQ2_XXS, main branch

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
512	128	0	4.793	106.82	5.452	23.48
512	128	512	4.984	102.73	6.375	20.08
512	128	1024	5.357	95.58	6.191	20.68
512	128	1536	5.062	101.15	6.290	20.35
512	128	2048	5.168	99.07	6.559	19.51

IQ2_XXS_R4

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
512	128	0	3.467	147.69	5.508	23.24
512	128	512	3.764	136.03	5.964	21.46
512	128	1024	3.573	143.31	6.292	20.34
512	128	1536	3.660	139.88	6.341	20.19
512	128	2048	3.729	137.29	6.620	19.33

IQ2_XXS, PR

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
512	128	0	1.778	288.03	5.484	23.34
512	128	512	1.860	275.28	5.685	22.52
512	128	1024	1.948	262.82	5.848	21.89
512	128	1536	2.040	250.93	6.158	20.78
512	128	2048	2.131	240.32	6.322	20.25