Fix train_mlp_nerf and save the model at the end of training

[97littleleaf11]

The near and far parameter of render_image during evaluation is not consistent with training stage. Currently this is not a real problem since there is a default near and far value in ray_marching for unbound scenes.

Other modifications:

• reuse the MIPNERF360_UNBOUNDED_SCENES and NERF_SYNTHETIC_SCENES.
• save the model at the end of the training. Since it still takes some time for MLP to converge, it would be annoying for first-time user who forgot to generate results.

[liruilong940607]

Thanks for the cleanup!!