More efficient way of allocating voxel hash blocks in context of dynamic surfaces

[Algomorph]

This is an optimization of #6

• Project corners of forward-warped activation candidate blocks onto depth image
• Bin projected candidate blocks into BxB rectangles on the depth image by recording the depth range (two floats) of each forward-warped candidate block into each BxB bin. I suggest 4x4 pixels for the bin size, but this should be parameterized. Probably, need to experiment how many blocks tend to fall into each bin first to determine optimal range array size for each bin. Overflow should be ignored.
• Query pixel values for each block: if depth is within truncation distance + depth range of recorded block depth range, mark block to be allocated.

This reduces complexity from O(block_count * depth_sample_count) to O(bin_size * depth_sample_count) + O(block_count).