ggml-cpu: add Q4_0 8x8 gemv/gemm for riscv vlenb=16 case

[hongyang-7]

Summary

This PR extends existing riscv gemv/gemm kernels for q4_0_q8_0 in 8x8 layout for vlenb=16.

Key Changes

• Enable rvv gemv/gemm kernel for q4_0_q8_0 in 8x8 layout, which is not in real use in master branch, despite the 8x8 kernel implementation (vlenb>=32) already provided.
• Make the 8x8 kernel support vlenb=16 case, which is the choice of current mainstream high-performance hardware like SOPHGO2044.

Perplexity Test

./bin/llama-perplexity -m models/Qwen3.5-0.8B-Q4_0.gguf -f path/to/test_prompt.txt

model	master PPL	opt PPL
Qwen3.5-0.8B-Q4_0	17.8834 +/- 0.83491	17.8859 +/- 0.83492

Performance Test

Performance is tested on a real high-performance RISC-V server.

Env	Value
CPU	SG2044 64core riscv processor 2.6GHz
OS	openEuler 24.03 (LTS)
kernel	Linux 6.12
gcc	gcc-15.3.0
bench tool	numactl -C 0-15 -m 0 llama-batched-bench

Llama-3.2-1B-Instruct-Q4_0.gguf (738M)

Batch size	PP t/s (master)	PP t/s (opt)	PP boost (opt/master)	TG t/s (master)	TG t/s (opt)	TG boost (opt/master)
1	21.00	90.46	431%	15.54	20.04	129%
2	21.07	91.38	434%	16.81	29.70	177%
4	21.10	91.41	433%	17.37	48.91	282%
8	21.13	92.05	436%	17.69	53.80	304%

PP: 4.3x speedup
TG: 1.2x~3.0x speedup

Meta-Llama-3-8B-Instruct.Q4_0.gguf (4.4G)

Batch size	PP t/s (master)	PP t/s (opt)	PP boost (opt/master)	TG t/s (master)	TG t/s (opt)	TG boost (opt/master)
1	2.89	13.63	472%	2.40	3.50	146%
2	2.90	13.67	471%	2.64	5.23	198%
4	2.90	13.65	471%	2.68	10.16	379%
8	2.90	13.65	471%	2.71	10.93	403%

PP: 4.7x speedup
TG: 1.4x~4.0x speedup

Future Work

• Optimize other layout for more quantization types.
• Try to remove #if defined __riscv_zvfh path from top ggml-cpu/repack.cpp

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[ggml-gh-bot]

Hi @hongyang-7, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[hongyang-7]

Add Requirements section to respect PR Template.