metal : fix q5_k mul_mv register spill

[ggerganov]

cont #20398

Noticed too high register pressure in the q5_k vec kernel:

METAL_CAPTURE_ENABLED=1 GGML_METAL_CAPTURE_COMPUTE=8 ./bin/llama-completion -m ~/models/qwen2.5-3b-coder/ggml-model-q5_k.gguf -fa 1 -p "hello" -n 10 -no-cnv --top-k 1

Before:

After:

Perf:

Model	Test	t/s master	t/s gg/metal-mul-mv-q5_k-spill	Speedup
gemma2 9B Q5_K_M	pp1	56.27	61.67	1.10
gemma2 9B Q5_K_M	pp2	74.87	80.06	1.07
gemma2 9B Q5_K_M	pp3	83.92	89.26	1.06
gemma2 9B Q5_K_M	pp4	128.29	127.65	0.99
gemma2 9B Q5_K_M	tg32	57.22	62.00	1.08
gemma3 4B Q5_K_M	pp1	99.76	109.82	1.10
gemma3 4B Q5_K_M	pp2	146.22	159.29	1.09
gemma3 4B Q5_K_M	pp3	173.88	185.46	1.07
gemma3 4B Q5_K_M	pp4	236.71	236.36	1.00
gemma3 4B Q5_K_M	tg32	103.52	114.42	1.11
llama 8B Q5_K_M	pp1	78.57	84.24	1.07
llama 8B Q5_K_M	pp2	100.28	104.78	1.04
llama 8B Q5_K_M	pp3	112.15	114.89	1.02
llama 8B Q5_K_M	pp4	166.52	166.64	1.00
llama 8B Q5_K_M	tg32	79.30	84.32	1.06
qwen2 3B Q5_K_M	pp1	127.54	136.71	1.07
qwen2 3B Q5_K_M	pp2	183.02	196.62	1.07
qwen2 3B Q5_K_M	pp3	216.24	226.41	1.05
qwen2 3B Q5_K_M	pp4	276.70	276.13	1.00
qwen2 3B Q5_K_M	tg32	128.72	140.52	1.09
qwen2 7B Q5_K_M	pp1	79.31	86.19	1.09
qwen2 7B Q5_K_M	pp2	100.62	107.88	1.07
qwen2 7B Q5_K_M	pp3	110.58	117.67	1.06
qwen2 7B Q5_K_M	pp4	179.16	179.97	1.00
qwen2 7B Q5_K_M	tg32	79.47	86.58	1.09
qwen3 0.6B Q5_K_M	pp1	266.12	292.98	1.10
qwen3 0.6B Q5_K_M	pp2	483.89	503.42	1.04
qwen3 0.6B Q5_K_M	pp3	634.80	667.10	1.05
qwen3 0.6B Q5_K_M	pp4	733.00	733.70	1.00
qwen3 0.6B Q5_K_M	tg32	280.87	300.45	1.07