Add EXPERIMENTAL int8 graph-compressed bundle (int8/)

embeddinggemma-300m Core AI fp32 static (seq 256), rebuilt on Linux box 2026-08-17; producer coreai-core 1.0.0b2; tokenizer bit-identical to mlboydaisuke@d9a60a18, main.mlirb +262 B (conversion is not byte-deterministic)

initial commit