Benchmark to make sure `groupby_indices` is necessary

[dafeda]

The idea of groupby_indices was put together while trying to make Troll run in reasonable time. Many things have changed since, and my benchmarking while working with Troll was not rigorous. This trick introduces some complexity and we should make sure it's worth it. I therefore think we should try running at least Drogon, and perhaps even Troll with and without it.

https://github.com/equinor/iterative_ensemble_smoother/blob/cf2c87270d31cbc1539e9d9cfb64f3167333233c/src/iterative_ensemble_smoother/experimental.py#L281-L295

• Trick should be most beneficial when parameters are highly correlated, i.e., a nice and smooth gaussian field
• Perhaps the trick provides more benefit in later iterations because updated parameters are more correlated? Speculative

[dafeda]

Ran with and without trick on a gaussian field with 1 million parameters, 50 observations and 25 realizations. The sigma=10 controls the smoothness of the field. It runs faster without the trick.

    p = 1 mill, m = 50, N = 25, sigma = 10
    With trick:    1.8s, 1.7s, 1.9s (Mean)
    Without trick: 1.3s, 1.2s, 1.3s (Mean)

I also ran Drogon with and without the trick and it runs much faster without the trick My conclusion for now is that the trick is faster when the field is very smooth which it does not seem to be in practice (Drogon). So, I suggest we remove it.