Fix errors with metric accumulation

[sjmonson]

Summary

Fixes a issue in metric calculation that caused incorrect statistics at extreme changes in concurrency and an issue where the first decode token was not counted in total tokens per second.

Details

• Fixed issue where merged concurrency change events would double-count concurrency
• Ensure first decode token is counted when calculating total tokens per second

Test Plan

• Run unit tests: tox -e test-unit -- -m "regression and sanity"

---

• "I certify that all code in this PR is my own, except as noted below."

Use of AI

• Includes AI-assisted code completion
• Includes code generated by an AI application
• Includes AI-generated tests (NOTE: AI written tests should have a docstring that includes ## WRITTEN BY AI ##)

[Copilot]

Pull Request Overview

This PR fixes two metric calculation issues in the benchmark statistics system: a double-counting bug in concurrency calculations when events are merged due to epsilon tolerance, and incorrect token-per-second calculations that excluded the first decode token.

• Fixed concurrency metric accumulation logic to prevent double-counting when events are merged
• Corrected token-per-second calculations to include the first decode token by adding 1 to prompt token counts
• Added comprehensive regression and edge case tests for the concurrency calculation fixes

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
src/guidellm/objects/statistics.py	Restructured concurrency event processing to fix double-counting in merged events
src/guidellm/benchmark/benchmark.py	Added +1 to prompt tokens to include first decode token in total token calculations
tests/unit/objects/test_statistics.py	Added regression tests for concurrency double-counting and epsilon edge cases