v2: fused experts (4-bit) + 4-bit embeddings, even ~3.84GB shards (FFAI ffai convert)

4-bit (affine, group_size 64) FFAI quant of google/diffusiongemma-26B-A4B-it

initial commit