metal : add FA-vec kernels for head size 96

[ggerganov]

fix #12948

Increase tg performance at long contexts for models such as Phi-3 that have head size of 96.

./bin/llama-batched-bench -m ../models/phi-3-mini-128k-instruct/ggml-model-q8_0.gguf -c 32768 -b 4096 -ub 4096 -npp 0,512,4096,8192,16384 -ntg 128 -npl 1 -lv 1 -fa

• master

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
0	128	1	128	0.898	0.00	1.400	91.42	2.298	55.69
512	128	1	640	0.396	1294.24	1.510	84.74	1.906	335.77
4096	128	1	4224	2.173	1884.70	2.266	56.49	4.439	951.49
8192	128	1	8320	4.863	1684.49	3.128	40.92	7.991	1041.18
16384	128	1	16512	12.512	1309.50	4.867	26.30	17.378	950.15

• PR

PP	TG	B	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s	T s	S t/s
0	128	1	128	0.968	0.00	1.294	98.92	2.262	56.59
512	128	1	640	0.421	1217.31	1.345	95.19	1.765	362.55
4096	128	1	4224	2.150	1904.81	1.686	75.91	3.837	1100.97
8192	128	1	8320	4.848	1689.73	2.074	61.72	6.922	1201.97
16384	128	1	16512	12.522	1308.46	2.864	44.70	15.385	1073.24