Improve DR robustness

[mhubii]

Replaces #53.

Mainly addresses:

• Differentiable rendering robustness #36

Also addresses:

• Validate parse_data #72
• Default to visual meshes in CLI #70

Note to self: Simple loss

loss = torch.nn.functional.mse_loss(
    foreground_distance_maps, renders["camera"]
) 

converges better (given good initialization) than

loss = torch.nn.functional.mse_loss(
    foreground_distance_maps, renders["camera"]
) - torch.nn.functional.mse_loss(background_distance_maps, renders["camera"])