UPSTREAM PR #18401: ggml: add ggml_rope_comp

[loci-dev]

Mirrored from ggml-org/llama.cpp#18401

For the motivation of this proposal, please refer to this discussion: ggml-org/ggml#1401

For demo purposes, only 2 backends are implemented in this PR:

• CPU
• Metal

Main differences

ggml_rope_ext	ggml_rope_comp
One single API call	Multiple composable calls
High-level input params like freq_base, YaRN scaling	Lower-level input params, allow most of the static params to be customized by user code
Multiple kernels, one per mode	One single kernel, templated by mode (mrope/vision) or controlled via input arg (ordering neox/normal)
I32 type for position	F32 type for position; shape = 1D for text and 4D for m-rope
m-rope only supports neox ordering	m-rope now supports both neox and normal
Does not support offset n_rot	Allow offset n_rot [1]

[1] This is necessary because we may want to implement 2D-rope (vision mode) as 2 separated calls to ggml_rope_comp, one call with offset = 0 and the other with offset = n_rot/2. This is particularly useful for vision models like Pixtral, where the 2 parts of 2D-rope does not use the same freq configuration.

Performance

I still couldn't get a meaningful perf result due to --output csv not working with test-backend-ops. But at a glance via --output console, it provide the same performance as the existing ggml_rope_ext