Train-to-Test scaling explained: How to optimize your end-to-end AI compute budget for inference