Update benchmark_serving.py

[puneithk]

We need to have input and output token length for performance tracking.

---

Note

Low Risk
Single change to which JSON fields are dropped on save; no runtime serving or benchmark execution behavior changes.

Overview
Keeps per-request token lengths in default benchmark JSON exports so performance tracking can use them without turning on full detailed dumps.

When --save-result is used and --save-detailed is not, the saver still strips high-volume fields (ttfts, itls, generated_texts, errors) to limit file size, but it no longer removes input_lens and output_lens from the merged result.

^{Reviewed by Cursor Bugbot for commit 7a260a1. Bugbot is set up for automated code reviews on this repo. Configure here.}