metal : optimize dequant q6_K kernel

[akretz]

In this PR I have optimized the Metal q6_K dequantization kernel. The basic idea is to load the weights into 32 bit variables and to do the bit twiddling on them. That gives a total 4-5% speed improvement on prompt processing on my M1 Max.

Model	Test	t/s master	t/s dequant_q6_metal	Speedup
llama 8B Q6_K	pp64	289.95	303.06	1.05
llama 8B Q6_K	pp128	320.58	335.25	1.05
llama 8B Q6_K	pp256	332.92	347.99	1.05
llama 8B Q6_K	pp512	339.55	354.55	1.04
llama 8B Q6_K	pp1024	337.06	351.88	1.04
llama 8B Q6_K	pp2048	330.56	344.86	1.04
qwen2 14B Q6_K	pp64	153.16	159.54	1.04
qwen2 14B Q6_K	pp128	169.42	176.74	1.04
qwen2 14B Q6_K	pp256	177.42	185.07	1.04
qwen2 14B Q6_K	pp512	179.47	187.02	1.04
qwen2 14B Q6_K	pp1024	178.11	185.53	1.04
qwen2 14B Q6_K	pp2048	174.79	181.95	1.04

[ggerganov]

M2 Ultra results:

./scripts/compare-commits.sh master pr/11892 \
    -m models/llama-8b-v3/ggml-model-q6_k.gguf \
    -m models/llama-8b-v3/ggml-model-q4_k.gguf \
    -m models/qwen2.5-7b-coder/ggml-model-q6_k.gguf \
    -m models/qwen2.5-7b-coder/ggml-model-q4_k.gguf \
    -t 1 -fa 1 -p 1,2,3,4,8,16,32,64,128,512

Model	Test	t/s master	t/s pr/11892	Speedup
llama 8B Q4_K_M	pp1	90.67	90.47	1.00
llama 8B Q4_K_M	pp2	123.40	123.06	1.00
llama 8B Q4_K_M	pp3	142.98	143.06	1.00
llama 8B Q4_K_M	pp4	146.43	147.71	1.01
llama 8B Q4_K_M	pp8	188.94	184.88	0.98
llama 8B Q4_K_M	pp16	233.41	235.22	1.01
llama 8B Q4_K_M	pp32	460.10	463.28	1.01
llama 8B Q4_K_M	pp64	700.10	704.51	1.01
llama 8B Q4_K_M	pp128	922.30	930.33	1.01
llama 8B Q4_K_M	pp512	1135.04	1142.44	1.01
llama 8B Q4_K_M	tg128	90.01	90.07	1.00
llama 8B Q6_K	pp1	74.39	74.83	1.01
llama 8B Q6_K	pp2	97.57	97.98	1.00
llama 8B Q6_K	pp3	107.83	108.99	1.01
llama 8B Q6_K	pp4	143.78	153.44	1.07
llama 8B Q6_K	pp8	178.59	196.38	1.10
llama 8B Q6_K	pp16	221.91	232.37	1.05
llama 8B Q6_K	pp32	438.59	457.39	1.04
llama 8B Q6_K	pp64	668.77	695.76	1.04
llama 8B Q6_K	pp128	874.08	915.95	1.05
llama 8B Q6_K	pp512	1072.01	1123.14	1.05
llama 8B Q6_K	tg128	75.16	75.39	1.00
qwen2 7B Q4_K_M	pp1	91.65	91.24	1.00
qwen2 7B Q4_K_M	pp2	121.88	121.57	1.00
qwen2 7B Q4_K_M	pp3	141.70	141.80	1.00
qwen2 7B Q4_K_M	pp4	157.67	158.70	1.01
qwen2 7B Q4_K_M	pp8	198.63	200.83	1.01
qwen2 7B Q4_K_M	pp16	251.80	253.93	1.01
qwen2 7B Q4_K_M	pp32	496.30	500.68	1.01
qwen2 7B Q4_K_M	pp64	743.22	748.85	1.01
qwen2 7B Q4_K_M	pp128	960.58	966.90	1.01
qwen2 7B Q4_K_M	pp512	1226.83	1234.10	1.01
qwen2 7B Q4_K_M	tg128	90.80	90.53	1.00
qwen2 7B Q6_K	pp1	77.33	77.98	1.01
qwen2 7B Q6_K	pp2	101.18	101.86	1.01
qwen2 7B Q6_K	pp3	114.21	114.81	1.01
qwen2 7B Q6_K	pp4	153.18	165.03	1.08
qwen2 7B Q6_K	pp8	186.73	204.02	1.09
qwen2 7B Q6_K	pp16	238.37	249.68	1.05
qwen2 7B Q6_K	pp32	470.06	490.19	1.04
qwen2 7B Q6_K	pp64	705.50	736.11	1.04
qwen2 7B Q6_K	pp128	899.02	944.28	1.05
qwen2 7B Q6_K	pp512	1156.60	1211.02	1.05
qwen2 7B Q6_K	tg128	77.51	77.53	1.00

Included Q4_K as well since it has some Q6_K tensors - seems like a ~1% improvement there too.