feat: Enable pip caching to speed up execution

[wolfmagnate]

Description

Currently, execute.py launches a fresh ephemeral container (--rm) for each task. As a result, pip install rebuilds wheels from source for dependencies (e.g., numpy, grpcio) every single time. This compilation step dominates the execution time, making the evaluation extremely slow.

This PR mounts a local directory to /tmp/pip_cache inside the container to persist built wheels across tasks.

Impact

• Performance: Significantly reduces the installation time for subsequent tasks by reusing built wheels.
• Safety: Environment isolation is preserved. Only build artifacts (wheels) are shared, while installed packages (site-packages) remain isolated per container.