oxicuda-lm 0.2.0

Large language model inference primitives for OxiCUDA: BPE tokenizer, transformer layers with KV cache, GPT-2 and LLaMA architectures — pure Rust, zero CUDA SDK dependency.