ChatGPT as Linguistic Equalizer? Quantifying LLM-Driven Lexical Shifts in Academic Writing