benchmark execution issues

[jmatejcz]

Describe the bug
Investigate issues that sometimes occurs when running benchmarks:

• Model (qwen3:14b for example) sometimes gets stuck on thinking till timeout
• When running long manipulation benchmark with large models that take a lot of RAM, the python process after couple hours, gets killed because of RAM usage, which suggests that there is some memory leakage.
• When a large model takes up a lot of VRAM, there is no space for openset models and program thrors CUDA OOM, should we adress that issue?

To Reproduce

1. Run manipulation benchmark with qwen3:14b, no thinking
2. Run manipulation benchmark with large model like 32GB on CPU
3. Run manipulation benchmark with large model like 32GB on GPU

Expected behavior
A clear and concise description of what you expected to happen.

Screenshots
If applicable, add screenshots to help explain your problem.

Platform

• OS: Ubuntu 22.04
• ROS 2 Humble

Version
Release number or commit hash.

Additional context
Add any other context about the problem here.