Some optimize and build warning fix for LoongArch

[MQ-mengqing]

• ggml : optimize convert f32<->f16 for loongarch_asx
• ggml : optimize loongarch_asx extend i16,i8,u8 to i32,i16
• ggml : Fix warnings when run cpu CI locally on LoongArch

[ggerganov]

cc @junchao-loongson

[junchao-loongson]

perf record -e cache-misses ./build/bin/llama-cli -m ../qwen2-1_5b-instruct-q3_k_m.gguf -p "I believe the meaning of life is" -n 128

Before:
  26.02%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_q3_K_q8_K
  21.30%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_f16
  20.03%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_q4_K_q8_K
  10.55%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_q6_K_q8_K  

After:
  33.47%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_q3_K_q8_K
  25.05%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_q4_K_q8_K
  11.67%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_q6_K_q8_K
   5.35%  llama-cli  libc.so.6                      [.] memcpy
   4.63%  llama-cli  libggml-cpu.so                 [.] ggml_compute_forward_mul_mat
   2.94%  llama-cli  llama-cli                      [.] common_sampler_sample(common_sampler*, llama_context*, int, bool)
   2.19%  llama-cli  libc.so.6                      [.] memset
   1.89%  llama-cli  libggml-cpu.so                 [.] ggml_vec_dot_f16

The ggml_vec_dot_f16 hotspot function is clearly mitigated

Before:
> $ ./build/bin/llama-bench -m ../qwen2-1_5b-instruct-q3_k_m.gguf                                                                                                                                                                                   [±master]
| model                          |       size |     params | backend    | threads |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
| qwen2 1.5B Q3_K - Medium       | 780.32 MiB |     1.54 B | CPU        |       8 |         pp512 |         21.40 ± 0.01 |
| qwen2 1.5B Q3_K - Medium       | 780.32 MiB |     1.54 B | CPU        |       8 |         tg128 |         20.91 ± 0.07 |

build: c3db0480 (4645)

After:
> $ ./build/bin/llama-bench -m ../qwen2-1_5b-instruct-q3_k_m.gguf                                                                                                                                                                                  [±pr11709]
| model                          |       size |     params | backend    | threads |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
| qwen2 1.5B Q3_K - Medium       | 780.32 MiB |     1.54 B | CPU        |       8 |         pp512 |         23.09 ± 0.07 |
| qwen2 1.5B Q3_K - Medium       | 780.32 MiB |     1.54 B | CPU        |       8 |         tg128 |         21.44 ± 0.09 |

build: 99bbe263 (4656)

Benchmark has a small improvement

LGTM！