CUDA-MSST-Infer

refactor: 将Web UI从内联HTML提取到独立文件，优化可读性和文件选择按钮

feat: add runtime controls for CUDA Graph, model persistence, and precision mode; improve memory management and localize web UI

perf: add CUDA Graph caching for attention, optimize single-chunk mode and band processing

feat: enhance profiling capabilities with detailed timing statistics for attention and feedforward layers

perf: add profiling and runtime tuning controls