[WIP] Custom iq4_xs vulkan kernel

[pwomser]

Overview

This PR adds a custom kernel for the IQ4_XS quantization that increases token generation on unsloth/Qwen3.8-27B-UD-Q3_K_XL from ~36 t/s to ~37.2 t/s on my machine. The initial version was created by ShinkaEvolve with Qwen3.8-27b as LLM for the evolution and then reviewed and cleaned up by me.

Additional information

Benchmark on unsloth/Qwen3.8-27B-UD-Q3_K_XL (IQ4_XS takes up ~40% of total compute on that model) on my RX 9070 XT (radv):
./build/bin/llama-bench -m Qwen3.8-27B-UD-Q3_K_XL.gguf -ngl 99 -fa on -n 3 -r 4096 -p 0

Old:

model	size	params	backend	ngl	fa	test	t/s
qwen35 27B Q3_K - Large	12.23 GiB	27.32 B	Vulkan	99	1	tg3	36.03 ± 0.08

New:

model	size	params	backend	ngl	fa	test	t/s
qwen35 27B Q3_K - Large	12.23 GiB	27.32 B	Vulkan	99	1	tg3	37.19 ± 0.08

Perplexity on wikitext-2-raw didn't change:

lama-perplexity -m Qwen3.8-27B-UD-Q3_K_XL.gguf -c 512 --chunks 30 -ngl 99 --seed 1234:

	PPL OLD	PPL NEW
wiki.valid.raw	PPL = 5.5262 +/- 0.15001	PPL = 5.5262 +/- 0.15001
wiki.train.raw	PPL = 8.2900 +/- 0.24976	PPL = 8.2900 +/- 0.24976
wiki.test.raw	PPL = 5.7448 +/- 0.15760	PPL = 5.7448 +/- 0.15760

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - Kernel was initially created with ShinkaEvolve with Qwen3.8-27b. Qwen3.8-27b and GPT5.6 were also used to review it before submission.

[pwomser]

Closing in favor of #28426