[CausalLM] optimize gpt-oss-20b performance

[EunjuYang]

see also : #3444

• In [CausalLM] support gpt-oss-20b #3444, we implemented CausalLM for gpt-oss-20b.
• It runs with Q4_0-FP32.
• We aims to optimize its speed and reduce its memory on Android.
• last update : 2025-09-15 11:00 KST by ejyang

WTD

• [Feature] Activate Q4_0-FP16 support in nntrainer. (08cc7ba)
• [Feature] Support Q4_0 type in embedding layer. (b2faa4d, [CausalLM] enable Q4_0 in Embedding Layer #3476 )
• [Op] FP16 compute functions
  • (done) FP16 frequency values : 000eb2e
  • (done) FP16 RMS norm : 6ed672e
  • (done) FP16 ggml GEMMs : e9b667d
    • Further optimization suggestion : use internal fp16 / int16 values in ggml kernel
  • (nyi) FP16 clamp
• [Feature] Create Q4_0-FP16 weight bin file for gpt-oss-20b
• [code_clean] Update moe-cached.cpp to inherit moe.cpp
• [Test] Q4_0-FP16 for qwen3-30b / gpt-oss-20b
• [Improvement] moe-cached layer profiling
• [Test] Apply KV-cache loading [CausalLM] Sync gpt-oss with kv-cache-update #3490
• [Improvement] update gpt-moe to process expert with MM for prefill [CausalLM] updeate gpt-moe-cached to support MM #3478

Log

• (2025-09-10) Q4_0 embedding layer is not valid (accuracy drop is significant)