segmentation fault when using `open3d.geometry.PointCloud.hidden_point_removal`

**Describe the bug**
I use `open3d` and `hidden_point_removal` for months without any issue.
Recently, I got one case which causes me troubles.

A segmentation fault occurs when using `hidden_point_removal` function and I didn't succeed to catch why. This happened only once, but it is still reproductible.

**To Reproduce**
Steps to reproduce the behavior:

Here are the txt files containing datasets:
- [test_bad.txt](https://github.com/intel-isl/Open3D/files/5891253/test_bad.txt)
- [test_good.txt](https://github.com/intel-isl/Open3D/files/5891254/test_good.txt)

and a small script to reproduce the issue. It will load the data from both txt files, check convex hull, and then use hidden_point_removal function for each dataset.
```py
import numpy as np
import open3d as o3d

def load_bad_data():
xyz = np.loadtxt("test_bad.txt", delimiter=",")

assert xyz.shape == (157862, 3)

min_bound = np.min(xyz, axis=0)
max_bound = np.max(xyz, axis=0)
assert np.allclose(min_bound, np.array([-48.94669787, -40.96437845, -3.41310365]))
assert np.allclose(max_bound, np.array([50.92755409, 59.71741356, 17.19930483]))

bound_diff = max_bound - min_bound
assert np.allclose(bound_diff, np.array([99.87425196, 100.68179201, 20.61240848]))

return xyz

def load_good_data():
xyz = np.loadtxt("test_good.txt", delimiter=",")

assert xyz.shape == (1172, 3)

min_bound = np.min(xyz, axis=0)
max_bound = np.max(xyz, axis=0)
assert np.allclose(min_bound, np.array([-58.16689297, -50.24060715, -4.05097986]))
assert np.allclose(max_bound, np.array([49.4774204, 51.00874277, 5.42619586]))

bound_diff = max_bound - min_bound
assert np.allclose(bound_diff, np.array([107.64431336, 101.24934993, 9.47717572]))

return xyz

def find_visible_points(xyz):
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(xyz)

diameter = np.linalg.norm(
np.asarray(pcd.get_max_bound()) - np.asarray(pcd.get_min_bound())
)
radius = diameter * 50

# 🔽 something wrong happens here
_, visible_index = pcd.hidden_point_removal([0, 0, 0], radius)

return visible_index

def display_pointcloud(xyz):
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(xyz)

hull, _ = pcd.compute_convex_hull()
hull_ls = o3d.geometry.LineSet.create_from_triangle_mesh(hull)
hull_ls.paint_uniform_color((1, 0, 0))
o3d.visualization.draw_geometries([pcd, hull_ls])

if __name__ == "__main__":
# happy test
xyz_good = load_good_data()
display_pointcloud(xyz_good)
find_visible_points(xyz_good)

# when the bug occurs
xyz_bad = load_bad_data()
display_pointcloud(xyz_bad) # this shows that the bug is not related to the convex hull computation
find_visible_points(xyz_bad) # this will generate a segmentation fault

print("Yeah!") # this is never executed due to the segmentation fault

```

This will work as expected until the `find_visible_points(xyz_bad)` statement generates a segmentation fault:
```
$ python -X faulthandler test_open3d.py
Windows fatal exception: access violation

Current thread 0x00005eb0 (most recent call first):
File "test_open3d.py", line 46 in find_visible_points
File "test_open3d.py", line 68 in
Segmentation fault
```

**Expected behavior**
No segmentation fault, or at least raise an Exception python can catch.

**Environment:**

- Operating system: Windows 10 64-bit
- Python version: Python 3.8.5
- Open3D version: 0.12.0
- Is this remote workstation?: no
- How did you install Open3D?: from pip in a fresh conda environment

**Additional context**

I would be good to deal with this segmentation fault if Python allowed me to catch / handle the signal, but this completely shut off my script.

Hints:
- open3d related codebase : https://github.com/intel-isl/Open3D/blob/master/cpp/open3d/geometry/PointCloud.cpp#L587-L647
- python signal handler for segmentation fault : https://bugs.python.org/issue1215