Need post-run statistics

[FindHao]

We only have statistics per-op for now. Need a post-run analysis to summarize results across ops.

[xuzhao9]

run.py is specialized towards comparing kernels with the same set of inputs. If the goal is to compare/reduct in other dimensions (e.g., across precision), I tend to organize post-run analysis as individual benchmarks that uses Tritonbench as a library: https://github.com/pytorch-labs/tritonbench/tree/main/benchmarks