Add support for hysteresis thresholds for benchmarks

[chevdor]

Currently, considering calls for which the code does not change, we can see weights oscillating.
This is expected due to the nature of the benchmarks.

For instance 6_605_000 → 7_054_000 → 6_856_000.

We could implement configurable thresholds for the weight differences and handle them with an hysteresis to limit the noise.

If we set the threshold(s) to 0, we would get the current behavior where any change in weight is considered.

However, setting this threshold to 10% for instance, would not update the weights unless the change is over the threshold. As a result, we would greatly limit the noise in the weight updates.

We can use different thresholds for increases and decreasses.

This would make the diff and json smaller, and the reviews easier/faster as we could focus more on the changes that do matter.

[chevdor]

For some context, here is how a diff looks like today.