[Refactor] guidellm.backend.openai performance improvements

[markurtz]

Improve performance on top of the current rewrite to remove overheads such as too many async updates, too many decodes, and n^2 string chaining for gathering tokens