ggml : add mrope kernel for metal

[ngxson]

This gives x1.5 generation speed for Qwen VL models (tested on Macbook M3 Max)

master branch:

model	size	params	backend	threads	test	t/s
qwen2vl 7B Q8_0	7.54 GiB	7.62 B	Metal,BLAS	10	pp512	526.06 ± 6.63
qwen2vl 7B Q8_0	7.54 GiB	7.62 B	Metal,BLAS	10	tg128	22.83 ± 1.36

This PR:

model	size	params	backend	threads	test	t/s
qwen2vl 7B Q8_0	7.54 GiB	7.62 B	Metal,BLAS	10	pp512	604.64 ± 0.44
qwen2vl 7B Q8_0	7.54 GiB	7.62 B	Metal,BLAS	10	tg128	33.54 ± 0.04

---

Tested with test-backend-ops -o ROPE, result is OK