ggml-cuda: widen Q4_K MMVQ weight loads to global_load_b128

[Annieren]

Set VDR_Q4_K_Q8_1_MMVQ to 8 so each thread owns a full 32-byte Q4_K group and loads its weights as two int4 (lowered to global_load_b128), cutting weight-load instructions on the bandwidth-dominant side. Activations stay scalar b32 (block_q8_1 is 36B, not 16B-aligned). Numerically equivalent to the VDR=2 path. Measured +1-2% typical decode on the MMVQ path (gfx1151,GGML_CUDA_DQ_MMV=0), up to +6-7% on some models; Q4_K-only, no effect on the
default DQ path.

Q4_K MMVQ b128 sweep — DQ0 results

Decode throughput (tg128 t/s), baseline VDR=2 → new VDR=8

Model	Quant	DQ0 base	DQ0 new	DQ0 Δ%
gemma-4-E2B	Q4_K	93.36	99.69	+6.8
gemma-4-26B-A4B	Q4_K	44.51	47.39	+6.5
SmolLM2-1.7B	Q4_K	130.71	136.76	+4.6
Qwen2.5-3B	Q4_K	84.49	87.94	+4.1
Llama-2-7B	Q4_K	43.24	44.91	+3.9
gemma-4-31B	Q4_K	10.37	10.76	+3.7
Qwen2.5-7B	Q4_K	44.78	45.87	+2.4
Qwen3.5-9B	Q4_K	36.40	37.23	+2.3
Qwen3-30B-A3B	Q4_K	69.39	70.95	+2.2
Llama-3.1-8B	Q4_K	40.54	41.31	+1.9
Qwen3-1.7B	Q4_K	128.89	131.13	+1.7
Qwen3.5-4B	Q4_K	60.46	61.21	+1.2
gemma-4-12b	Q4_K	26.09	26.34	+1.0
Qwen3-4B	Q4_K	64.54	65.18	+1.0
Qwen3-8B	Q4_K	39.29	39.65	+0.9
Qwen3.6-35B	Q4_K_M	58.53	59.00	+0.8
Qwen3.5-35B-A3B	Q4_K	57.82	58.17	+0.6
gemma-2-2b	Q4_K	93.61	94.11	+0.5
Qwen3.6-35B	Q4_K_XL	56.75	57.03	+0.5
Qwen2.5-0.5B	Q4_K	303.48	304.03	+0.2
DeepSeek-R1-Qwen-7B	Q4_K	44.75	44.82	+0.2
Qwen3.5-4B	Q4_0	65.42	65.41	+0.0
Qwen3.5-9B	Q4_0	39.37	38.90	+0.0
Qwen3.6-35B	Q5_K_XL	55.42	55.11	+0.0