[nlp] Optimize prompt payload size to reduce generation latency

[ambadi565]

• Prune unused context tokens to accelerate LLM inference speeds.