MCMC Strategy does not respect cap_max if the number of gaussians exceeds the cap already

[vincentwoo]

This is a bit tricky because when initializing from SFM you may have altogether too many points, and would like to encourage MCMC to prune down to a target number. This would require adding a conditional branch in the MCMC strategy to handle this case (currently MCMC assumes it is always either at or below this cap).

(see: https://github.com/nerfstudio-project/gsplat/blob/main/gsplat/strategy/mcmc.py#L176)