ggml-cpu : add LoongArch64 LASX Q4_0 repack path

[ztsubaki]

Overview

LongArch has a LASX SIMD ISA similar to AVX2, which can be accelerated by repack. Previously, ggml-cpu only used LASX in the quantized dot product kernel on LoongArch. This PR adds a LASX kernel for the Q4_0_8x8_Q8_0 repack path for Q4_0 quantized models, with performance far exceeding the original dot product implementation and slightly higher than the OpenBLAS backend.

Additional information

quantize_mat_q8_0_4x8 is edited from x86 AVX2 implementation

Benchmark using Qwen 3.5 0.8B Q4_0 from unsloth on Loongson 3A6000 (4 core LA664)

Before(bec4772):

| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen35 0.8B Q4_0               | 473.21 MiB |   752.39 M | CPU        |       4 |           pp512 |         79.59 ± 0.03 |
| qwen35 0.8B Q4_0               | 473.21 MiB |   752.39 M | CPU        |       4 |           tg128 |         37.44 ± 0.34 |

Current submit:

| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen35 0.8B Q4_0               | 473.21 MiB |   752.39 M | CPU        |       4 |           pp512 |        132.13 ± 0.05 |
| qwen35 0.8B Q4_0               | 473.21 MiB |   752.39 M | CPU        |       4 |           tg128 |         40.16 ± 0.37 |

OpenBLAS backend:

| model                          |       size |     params | backend    | threads |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| qwen35 0.8B Q4_0               | 473.21 MiB |   752.39 M | BLAS       |       4 |           pp512 |        131.72 ± 3.65 |
| qwen35 0.8B Q4_0               | 473.21 MiB |   752.39 M | BLAS       |       4 |           tg128 |         36.27 ± 0.30 |

Compared to before, pp512 has a 1.66x speedup, tg128 has a 1.07x speedup, and compared to OpenBLAS, pp512 is almost the same, while tg128 has a 1.11x speedup.

The perplexity was verified using the top 10% of the wiki-test2 test set on the same model and the same CPU.

CPU backend(using new LASX repack path):

Final estimate: PPL = 20.6118 +/- 0.46939

OpenBLAS backend:

Final estimate: PPL = 20.5799 +/- 0.46865

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

Use AI to read and review the code, but did not use AI to generate the code.