Upload faithful 2-layer Kimi-K2.6 (bf16, .layers., untied lm_head)

Clear old 4-layer contents

Use native kimi_k25/deepseek_v3 modeling: drop auto_map + custom modeling files

Add tiny 4-layer random-weight Kimi-K2.6 for conversion testing

initial commit