This PR addresses three main points:

1. Adds functionality to compute (geometric mean) speedups between two submissions. Across all base workloads, it computes the speedup (runtime_1/runtime_2) and takes the geometric mean across all. If a submission does not reach the target, the time is replaced with max_runtime_budget + 1 second (as if it would have hit the target just after).
2. Fixes the scoring code to ignore runtimes on workload variants if the base workload couldn't be trained successfully. This addresses "To determine the fastest submission on a held-out workload, we only consider submissions that reached the target on the corresponding fixed workload. This protects us against extremely fast submissions that only work on a specific held-out workload and are useless as general algorithms." in point 4 of https://github.com/mlcommons/algorithmic-efficiency/blob/main/DOCUMENTATION.md#using-held-out-workloads-in-scoring. It did not affect the results for the given logs.
3. Fixes the max_tau to 4.0, following our rules: https://github.com/mlcommons/algorithmic-efficiency/blob/main/DOCUMENTATION.md#integrating-performance-profiles-for-the-benchmark-score