feat: Return unexpected kubectl-ai errors in RunOnce mode, display them in k8s-bench

[noahlwest]

• kubectl-ai returns unexpected errors encountered in kubectl-ai RunOnce mode
• k8s-bench run now counts unexpected errors as errors instead of failures
• k8s-bench analyze now accounts for errors encountered

Tested this with local served gpt-oss-20b, errors like 404, 400, tool parsing errors, etc. are saved and recorded as errors now in k8s-bench. Easiest way to reliably recreate errors is to use start a vllm instance and point to a model that doesn't exist, or use a model with known issues (like gpt-oss-20b on completions endpoint.)

• When unexpected error is encountered and returned to k8s-bench, we print usage/help text for kubectl-ai. This adds a little bit of noise to the logs but we can fix that in a later pr

[droot]

Thank you so much for quick turn around on this one. This addresses a critical issues.