[filters] Increase the volumetric capacity of voxel_grid

[markhedleyjones]

The current implementation of voxel_grid is too limited in its processing volume.
This is due to its use of an int32 to index voxels internally.
With a leaf size of 3cm (in each axis), it can only support a volume of 38.7m^3.

What's worse is that voxel_grid will still process your point-cloud even if that index variable overflows.
In this case, the filter adds noise to the output.
The filter does produce a warning, but it's not clearly telling the user that it will introduce rubbish into the output.

Context

The pcl::voxel_filter is used in may projects and is used as part of many higher level algorithms.
Those algorithms suffer from the addition of noise or limited volumetric resolution that voxel_grid.
Allowing voxel_grid to handle either smaller leaf sizes or larger volumes would improve those algorithms.
TixiaoShan/LIO-SAM is one such example, where it's use is currently limited to outdoor cases.
With the ability to reduce the leaf size, that project is able to perform better indoors.

[markhedleyjones]

Example

With leaf sizes smaller than 0.001, the filter does produce the warning [pcl::VoxelGrid::applyFilter] Leaf size is too small for the input dataset. Integer indices would overflow. but still returns a PCD. That PCD can contain noisy points as a result of the voxel index overflowing. Examples of that noise are shown below.

The original pcd file can be downloaded from sourceforge. It is Rf14.pcd within the kitchen.zip file.