v0.2.0

This release delivers multiple major features:

• support block diffusion LLM (LLaDA2-preview and LLaDA2);
• add an optimized batch inference on block diffusion LLM;
• support long sequence generation;
• support native CUDA graph capture and management;
• add SGLang as a backend for block diffusion LLM inference;
• support FP8 quant for LLaDA2-preview and LLaDA2;
• add an experimental feature to support the integration of dInfer and SGLang;
• lm_eval support two benchmarks (gsm8k and mbpp) on LLaDA2-preview and LLaDA2