Fix benchmark docs and behavior: LightGBM does not support GPU-accelerated inference

[coderabbitai]

Summary

The benchmark suite added in #101 incorrectly treats LightGBM as supporting GPU-accelerated inference. As noted by @jameslamb in this review comment, LightGBM's OpenCL/GPU support is limited to training only — there is no GPU-accelerated inference in LightGBM.

Problem

• README.md in the benchmark suite documents a "GPU" mode for LightGBM inference that does not exist.
• The benchmark code may be running LightGBM inference on CPU even when --device=gpu is specified, making the comparison misleading.

Suggested Fix

• Update the README.md to clarify that LightGBM inference always runs on CPU, regardless of the --device flag.
• Update the benchmark runner (run.py) to reflect that when comparing nvforest GPU inference against LightGBM, the LightGBM side is CPU-only inference.
• Ensure the output CSV and summary statistics accurately label LightGBM results as CPU inference.

References

• PR Add comprehensive benchmark suite comparing nvforest against sklearn, XGBoost, and LightGBM native inference #101 (source of this issue)
• LightGBM GPU tree learner source (training only)

Reported by @csadorf.