feat: add options for brownian_tree

[vmobilis]

Addition for leejet#1899.

[wbruna]

UI-wise, I believe turning it into a decoupled, specific property would be better: say, "brownian_tree_rng=cpu" instead of "brownian_tree_cpu". Then, we could consider if having a "brownian_tree_rng" implies turning on "brownian_tree". Either way sounds fine to me, so it'd be up to leejet.

Code-wise, the name -> subclass mapping is replicating StableDiffusionGGML::get_rng functionality; we should probably have instead a virtual constructor somewhere (src/core/rng.hpp would be natural for the declaration, but would need to be implemented elsewhere to avoid running into a chicken-and-egg issue with the hierarchy; so probably util.h would be a better place), and convert the name first with str_to_rng_type.

About the functionality itself... to be honest, I'm not sure what would be the best approach. Note that the BrownianTreeNoiseSampler::bridge member also creates an internal generator, and is recursive 🙂 I would suggest just either passing the secondary rng type to the Brownian tree constructor, and let it do its own thing as it is right now, or (as fszontagh suggested) add a virtual clone() to the rng class and duplicate the original generator edit: I see you are already copying the generator, so it'd be enough to copy the main one on bridge as well..

[vmobilis]

@wbruna, then maybe this, as you suggested:

1. noise_sampler=brownian_tree_cpu
   vvv
   noise_sampler=brownian_tree,brownian_tree_rng=cpu (default is std);
2. str_to_rng_type() and get_rng() are reused in denoiser and are made visible globally for that;
3. rngs in BrownianTreeNoiseSampler() are copied independently; since RNG is polymorphous, the RNG->clone() is added for that.

I didn't test it yet, but it compiles.