This PR fixes multiple out-of-bound issues when only a small number of (or even 0) Gaussians are in the view:

• compute_cumulative_intersects should not access cum_tiles_hit, which has size 0 if no Gaussian is in the scene;
• Per-point kernels such as compute_cov2d_bounds_tensor should not be launched with 0 thread blocks, otherwise a CUDA invalid configuration exception will be thrown.
• tile_bins returned by get_tile_bin_edges_tensor now has size tile_bounds[0] * tile_bounds[1] instead of num_intersects (could be 0 if no visible Gaussian), which makes sure access to tile_bins[tile_id] in rasterization kernels guaranteed to stay in the range of tile_bins.

In addition, per @kerrj's suggestion #84 (comment), rasterization functions now directly return a tensor with all pixel values set to the background.

Closes #84.