Move camera optimization out of datamanager and parallelize dataloading

[kerrj]

WIP for removing camera optimization from datamanger and moving it to the model. This will potentially allow significant speedups in dataloading as several people have already noted, since the most time consuming step in dataloading is ray generation, which currently depends on gradients from the camera optimizer and thus cannot be parallelized.

Currently this

• Removes all camera optimization code from datamanager
• Adds camera optimization to nerfacto.py and refactors param groups, optimizer config accordingly

Things that need to happen:

• Think about how to merge this PR without breaking everyone's custom method configs which rely on the camera optimizer being inside the datamanager
• Think about proper inheritance for using camera optimizer (currently it sits in nerfacto.py, but it might be better to move it inside base_model.py)
• Parallelize all the ray generation code inside a proper dataloader

[ksnzh]

https://github.com/nerfstudio-project/nerfstudio/blob/justin/camera_opt_refactor/nerfstudio/pipelines/base_pipeline.py#L363 may be fixed to assert isinstance(self.datamanager, (VanillaDataManager, ParallelDataManager))

[maturk]

I think this is ready for merge. Tested with Lerf repository out of the box, and warnings appeared as they should without breaking the pipeline. Users would have to pip install nerfstudio once again after this update, since viser version has been bumped up.

[maturk]

pytest failed :/