Bug: inverse_sigmoid(1.0) produces Inf in densify_from_depth_propagation

[anton-brandl]

In scene/gaussian_model.py line 486, opacity is initialized as:

opacities = inverse_sigmoid(1.0 * torch.ones(...))                                                                                                                                                                                   

Since inverse_sigmoid(x) = log(x / (1-x)), when x=1.0 this produces Inf,
causing CUDA errors during training.

Fix: Use 0.999 instead of 1.0 to achieve near-full opacity without the singularity:
opacities = inverse_sigmoid(0.999 * torch.ones(...))