Restore orphaned moe_fp32 solutions (reference-scalar + ggml baseline)

Rewrite llama.cpp q8_0 reference-scalar solution signature, now compatible with bench/dataset/raw.py

Add kernel in different precision:

Add reference-scalar llama.cpp solutions (naive scalar baseline, 16 defs)