Speculators v0.5.0: DFlash support and online training

The v0.5.0 release of Speculators introduces DFlash support and improves online training capabilities for speculative decoding models, focusing on single-pass draft token generation and enhanced architectural features. Key updates include integration with vLLM for efficient training and inference, detailed documentation, and benchmarks showcasing the performance of the DFlash algorithm, particularly in reasoning and code generation tasks.