HIP: branch-free SWAR for __vsub4 / __vcmpne4 / __vcmpeq4

[SimonTeixidor]

Overview

ggml/src/ggml-cuda/vendors/hip.h emulates these 3 intrinsics with a loop, this PR
replaces those loops with a branch-free SWAR.

Additional information

__vsub4 was previously forwarded to __vsubss4, which saturates, and is now
a wrapping subtract instead. That should be fine for two reasons. First it
matches CUDA's __vsub4, and second it passes test-backend-ops -o MUL_MAT
(1288/1288).

End-to-end benchmark (Strix Halo / gfx1151, llama-batched-bench, S_TG)

The percentage after each quant is its share of decode weight bytes in the affected types.

npl	Qwen3.8-27B-UD-IQ4_XS (19%)	Qwen3.8-27B-AD-IQ2_S (86%)
4	+5.4%	+23.1%
6	+2.1%	+21.6%

The five types that use these intrinsics are IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ3_S.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: I noticed that certain quants were faster/slower than others, at the same file size. Claude Opus found this (and a few other) issue and helped me implement a fix.