[Feat] Added pruning strategy from https://arxiv.org/abs/2511.16980

[AshadowZ]

Gradient-Driven Natural Selection for Compact 3D Gaussian Splatting

Summary

This PR introduces a pruning mechanism based on "Gradient-Driven Natural Selection for Compact 3D Gaussian Splatting" (https://arxiv.org/abs/2511.16980). This new mechanism implements budget-aware pruning to represent scenes with a more compact set of Gaussians. Additionally, this PR removes the MCMC strategy and includes mild code refactoring

Usage

To utilize the pruning mechanism, please enable enable_natural_selection and set the final_budget. The densification strategy will grow the model to 2.5x the final_budget during iterations 0–15,000, and then gradually prune it down to the final_budget between iterations 15,000 and 23,000

Performance Results

This implementation largely follows the original. I was too lazy to generate new tables recently, so please refer to the results in the original paper—there shouldn't be any significant deviation. It is worth noting that I replaced the adaptive adjustment mechanism for opacity regularization weights used in the original implementation, though this shouldn't make a significant difference