Convolution3D: use dynamic schedule for OpenMP

Each iteration does a radius search, which does not take the same amount of time for each point. Specifying no schedule usually results in a static schedule.
Related to #5775

Benchmarks with table_scene_mug_stereo_textured.pcd (nan points removed before convolution) on Intel Core i7-9850H:

GCC:
threads | 1    | 2    | 3    | 4    | 5    | 6    |
before  | 2267 | 1725 | 1283 | 1039 |  863 |  744 |
dynamic | 2269 | 1155 |  795 |  611 |  497 |  427 |

MSVC 2022 (release configuration):
threads | 1    | 2    | 3    | 4    | 5    | 6    |
before  | 2400 | 1886 | 1478 | 1176 |  972 |  857 |
dynamic | 2501 | 1281 |  919 |  704 |  593 |  537 |

filters/include/pcl/filters/impl/convolution_3d.hpp

@@ -249,7 +249,8 @@ pcl::filters::Convolution3D<PointInT, PointOutT, KernelT>::convolve (PointCloudO
   default(none) \
   shared(output) \
   firstprivate(nn_indices, nn_distances) \
-  num_threads(threads_)
+  num_threads(threads_) \
+  schedule(dynamic, 64)
   for (std::int64_t point_idx = 0; point_idx < static_cast<std::int64_t> (surface_->size ()); ++point_idx)
   {
     const PointInT& point_in = surface_->points [point_idx];