Use FP32 VoxCPM2 text prefill for CJK quality

Add unified decoder weights (fp16)

Add unified prefill+token-step decoder graph (fp16, weight-deduplicated)

text_prefill INT8: symmetric activations (TRT EP compatible)

text_prefill INT8: switch to QDQ (TRT-friendly), drop MatMulNBits sidecar

LM graphs: fp16-weight/fp32-compute defaults + MatMulNBits int8 variants, replacing dynamic-INT8

token-step INT8: MatMulNBits weight-only (block=32, fp32 accum) replacing dynamic quant

text-prefill INT8 external data (MatMulNBits)

text-prefill INT8: MatMulNBits weight-only (block=32, fp32 accum) replacing dynamic quant

voxcpm2 ONNX (INT8 weights, FP32 activations) — real-weights export

initial commit