[feat CudaRansac] Do not use CUDA RNGs. Make number of initial points configurable

[true-real-michael]

This PR speeds up RANSAC by eliminating the need to initialize CUDA random number generators

N Clouds   Current (s)    New (s)
10          1.272783     1.205917
20          2.316050     2.098319
30          3.237002     2.969012
50          5.064633     4.742269
80          7.821288     7.056089
100        10.663958     9.202390

The current solution generates random numbers for each voxel and each hypothesis for this voxel independently. It requires initialisation for all RNGs.

We can replace that with generating random float numbers (same for all voxels but different for each hypothesis). Then each voxel would interpret the float numbers differently, converting them into the valid point indices for the voxel.

Another positive feature is that it allows user to configure the number of initial RANSAC points via a parameter for map_leaf_points_cuda_ransac which is impossible in the current solution (only using INITIAL_POINTS_NUMBER constant).