Vulkan: Tune Vulkan mmq int dot shader for performance

[0cc4m]

Retune the DP4A matmul shaders I added in #12135 to extract more performance from them after a bugfix reduced their performance in #12722. Here are my test results:

RTX 3090:

model	size	params	backend	ngl	test	t/s fp16	t/s int dot master	t/s int dot PR	t/s coopmat1	t/s coopmat2	t/s CUDA
llama 8B Q4_0	4.33 GiB	8.03 B	Vulkan	99	pp512	1025.95 ± 1.11	1928.35 ± 8.34	1925.12 ± 6.65	3138.60 ± 28.29	4247.01 ± 60.18	5069.38 ± 18.59
llama 8B Q8_0	7.95 GiB	8.03 B	Vulkan	99	pp512	1000.95 ± 2.89	1898.78 ± 4.69	1894.99 ± 7.38	2749.46 ± 3.65	4329.58 ± 16.31	4932.49 ± 14.11

AMD Radeon RX 6800 XT:

model	size	params	backend	ngl	test	t/s fp16	t/s int dot master	t/s int dot PR	t/s ROCm
llama 8B Q4_0	4.33 GiB	8.03 B	Vulkan	99	pp512	922.14 ± 1.04	1284.95 ± 2.76	1463.92 ± 1.74	1678.14 ± 2.28
llama 8B Q8_0	7.95 GiB	8.03 B	Vulkan	99	pp512	902.82 ± 0.80	1060.93 ± 0.74	1250.73 ± 1.23	1618.84 ± 1.25

AMD Radeon Pro VII:

model	size	params	backend	ngl	test	t/s fp16	t/s int dot master	t/s int dot PR	t/s ROCm
llama 8B Q4_0	4.33 GiB	8.03 B	Vulkan	99	pp512	313.00 ± 0.60	390.98 ± 1.52	588.09 ± 0.42	1012.39 ± 0.46
llama 8B Q8_0	7.95 GiB	8.03 B	Vulkan	99	pp512	309.01 ± 0.48	317.19 ± 0.91	501.33 ± 0.55	398.53 ± 0.06

Intel A770:

model	size	params	backend	ngl	test	t/s fp16	t/s int dot master	t/s int dot PR	t/s SYCL
llama 8B Q4_0	4.33 GiB	8.03 B	Vulkan	99	pp512	165.47 ± 0.14	508.05 ± 0.95	734.05 ± 1.60	917.26 ± 6.47
llama 8B Q8_0	7.95 GiB	8.03 B	Vulkan	99	pp512	157.58 ± 0.17	493.80 ± 0.76	678.78 ± 0.75	893.41 ± 4.35