Fix splatfacto crash in eval when using viewer

[aayushg55]

This addresses the issue in #3253 where splatfacto training could crash during the evaluation step when interacting with the viewer (e.g., with --vis viewer+wandb). The problem occurs because interacting with the viewer renders a new image in _render_img(), where model.eval() is called at the start and model.train() at the end, regardless of the initial model state (see nerfstudio/viewer/render_state_machine.py#L151). If this happens during the actual eval step of the model, it flips the mode to training during evaluation, causing errors.

Changes made:

• Modified eval_iteration() to use the train_lock, ensuring a consistent model state during evaluation

Testing

• Verified that splatfacto no longer crashes during evaluation with viewer interactions.

Closes #3253

[jb-ye]

It looks good to me, but I want to have @brentyi to take a second pass on this.

[brentyi]

Makes sense to me too!