LSETA PyTorch-native embedding model training framework with fused Triton kernels, sequence packing, and multi-GPU parallelism.