Background context: go/memory-infra

One problem we have with memory benchmarking in the lab is that they have to be short (to run on bots & bisectable). Therefore, if someones land a CL that create a leakage of a very small object, we have a very low chance of catching the leakage by monitoring the memory metrics as we do today. ( crbug.com/703735#c4 )

The proposal here is to use metrics that track objects allocation to detect leak. This would help amplify the regressions of a leakage of thousands objects whereas each object costs only a few bytes.