[Epic] Accurate performance tracking

Current

Currently LiquidCache run benchmark on every PR, it looks like this:

Future

While it is nice to have a rough idea of what's going on with the performance, the run-to-run variance has been quite large.

The variance comes from two sites:

1. Upstream system variance. E.g., DataFusion has some quite high run-to-run variance that is mysterious to me. I sometimes observe quite large variances when running DataFusion on my local machine. This should not happen, and we need to investigate this a lot more.
2. The testbed variance. We run benchmark on ubicloud-standard-4-ubuntu-2404 but it still has high variance, because it's not a dedicated runner.

While we could definitely directly address the above issues, it can take a while to make meaningful progress.

Instead, I'm thinking about tracking invariants that remain unchanged no matter what hardware we benchmark on:

1. Number of allocations and deallocations, total size of allocations, number of small allocations, number of large allocations.
2. Run benchmark with cachegrind, but this is super slow
3. Check non-time-based internal metrics

cc @alamb who is also looking at performance tracking

This is an epic and names no files or tests. Start by reviewing the existing benchmark-on-PR setup, then investigate the listed sources of variance and compare allocation metrics, cachegrind, and other non-time-based metrics. Done requires a decided, implementable approach for accurate performance tracking.