Two-level randomization (1-select population, 2-split across variants)

[manio143]

While reading the paper [1] I saw that there's a potential for an experiment re-run currently not exactly supported by Excos.

1. Let's take 10% of users into an experiment
2. 50/50 control and treatment
3. Run experiment
4. Now we want to use the same users, but re-randomize the variant assignment

Currently the first part would be achieved with a Feature config with two variants at [0;0.05) and [0.05;0.1). To change randomization we need to alter the salt, but that will re-randomize the entire population, choosing a different set of 10% users.

To address it we need a feature level population constraint.
I've already added variant-level salts to support GrowthBook, so they can be reused to execute re-randomization with the selected user group at a feature level.