feat: add w4afp8 on Hopper GPUs

[foreverrookie]

Hi, from novita.ai team.

• Algorithm compatible with https://huggingface.co/Barrrrry/DeepSeek-R1-W4AFP8, but uses a custom weight layout (see convert_fp8_to_int4 in tests/generators.py).

test perf(W4Afp8 vs FP8) on H200. nvcc: 12.9.

groups	m/grp	n	k	W4 us	W4 GB/s	FP8 us	FP8 GB/s	Speedup
8	16	4096	7168	48	2643	68	3460	1.42x
8	24	4096	7168	48	2658	68	3480	1.42x
8	32	4096	7168	48	2690	68	3486	1.42x
8	40	4096	7168	48	2724	69	3499	1.44x
8	48	4096	7168	48	2721	68	3516	1.42x
8	56	4096	7168	86	1543	76	3170	0.88x
8	64	4096	7168	85	1564	75	3239	0.88x
8	16	7168	2048	32	2036	40	2963	1.25x
8	24	7168	2048	32	2049	40	2981	1.25x
8	32	7168	2048	32	2095	40	3004	1.25x
8	40	7168	2048	32	2120	40	3027	1.25x
8	48	7168	2048	32	2161	40	3059	1.25x
8	56	7168	2048	31	2244	44	2843	1.42x
8	64	7168	2048	42	1666	45	2768	1.07x
16	16	4096	7168	91	2782	131	3613	1.44x
16	24	4096	7168	92	2787	131	3633	1.42x
16	32	4096	7168	91	2832	131	3634	1.44x
16	40	4096	7168	91	2838	131	3655	1.44x
16	48	4096	7168	91	2859	131	3669	1.44x
16	56	4096	7168	128	2051	140	3466	1.09x
16	64	4096	7168	128	2068	141	3458	1.10x
16	16	7168	2048	56	2305	71	3346	1.27x
16	24	7168	2048	56	2336	72	3368	1.29x
16	32	7168	2048	55	2411	71	3402	1.29x
16	40	7168	2048	56	2423	72	3434	1.29x
16	48	7168	2048	56	2467	72	3457	1.29x
16	56	7168	2048	67	2094	75	3334	1.12x
16	64	7168	2048	68	2077	79	3170	1.16x