Remove redundant w8_bf16 engines (int8 weights fold to bf16 at build → identical size/speed to the bf16 baseline)

Add sm_120 TRT engines (AOT SWA, wide profile) for the quantized tiers

sm_120 SAME-L: AOT (graph-capturable) engines + tensorRT/README.md (#6)

Rebuild medium DiT fp16-mixed engines (sm_90 + new sm_120): attention now fuses, 4.3x faster (#4)