CPU: Fix loongarch lsx compilation error

[junchao-loongson]

Using the -DGGML_LASX=OFF parameter during compilation results in the following compilation errors. This is due to some 128-SIMD code not being correctly included within the #if defined(__loongarch_asx)...#endif block, as well as missing explicit type conversions.
在编译时使用-DGGML_LASX=OFF参数，会出现以下编译错误，这是由于部分128-simd代码没有正确包含在#if defined(__loongarch_asx)...#endif中，以及缺少一些显示的类型转换

/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h: In function ‘__lsx_f16x4_load’:
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h:1055:12: error: incompatible types when returning type ‘__vector(2) long long int’ but ‘__m128’ {aka ‘__vector(4) float’} was expected
     return __lsx_vld(tmp, 0);
            ^~~~~~~~~
In file included from /home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/quants.c:5:
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h: In function ‘__lsx_f16x4_load’:
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h:1055:12: error: incompatible types when returning type ‘__vector(2) long long int’ but ‘__m128’ {aka ‘__vector(4) float’} was expected
     return __lsx_vld(tmp, 0);
            ^~~~~~~~~
In file included from /home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c:14:
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/vec.h: In function ‘ggml_vec_dot_f16_unroll’:
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/vec.h:236:67: error: incompatible types when initializing type ‘float’ using type ‘__vector(2) long long int’
         GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
                                                                   ^~~~~~~~~~~~~~~~~
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/vec.h:255:13: error: incompatible types when initializing type ‘__m128’ {aka ‘const __vector(4) float’} using type ‘__vector(2) long long int’
             GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
             ^~~~~~~~~~~~~~~~~~~
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/vec.h: In function ‘ggml_vec_mad_f32’:
/home/junchao/work/ai/llama.cpp/ggml/src/ggml-cpu/vec.h:370:27: error: incompatible types when initializing type ‘__m128’ {aka ‘__vector(4) float’} using type ‘__vector(2) long long int’
         GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
                           ^~~~~~~~~~~~~~~~~

• Added explicit type conversions to prevent compilation errors
• Moved functions implemented by lsx into the appropriate macros

Make sure to read the contributing guidelines before submitting a PR

[ggerganov]

We've improved support for adding self-hosted runners to the CI: https://github.com/ggml-org/llama.cpp/blob/master/ci/README.md

You can consider adding a loongarch runner to make things more stable in the future.