Describe the bug
#2533 was resolved stating that it had minimal dependencies (thank you!).

However, we wanted to update to open3d==0.10.0, but ran into #1937 (JVisualizer).
So then I wanted to update to open3d==0.12.0, but then ran into #2825 (dep on sklearn, which hasn't been released yet).

However, I don't understand why the base Open3D-ML/requirements.txt is merged into this Python package's requirements:
https://github.com/intel-isl/Open3D/blob/v0.12.0/python/setup.py#L52-L55

Why not use extras_require to use the ML dependencies, rather than doing a runtime-switch based on what packages are available? An example:
https://github.com/mikedh/trimesh/blob/4cacbffcd9bd1a637bf05b638bc3376464e3577b/setup.py

e.g.:

open3d  # Only Open3D deps, no ML stuff
open3d[ml]  # ML deps, but does not install PyTorch / TF
open3d[torch]  # ML deps + PyTorch
open3d[tensorflow]  # ML deps + TF
open3d[all]  # All the things!!!

To Reproduce
See this gist: https://gist.github.com/EricCousineau-TRI/ee119a3feeed77ac289ed06b01e587cc
Important snippet:

scikit-learn==0.24.1
    # via sklearn
...
sklearn==0.0
    # via open3d

(Note: sklearn isn't important, as I'd expect it to be fixed once a new release is cut; scikit-learn is the main thing here)

Expected behavior
I would have expected scikit-learn to not appear, or only appear of I installed a more explicit version.

Environment (please complete the following information):

• Operating system: Ubuntu 18.04
• Python version: Python 3.6
• Open3D version: 0.12.0
• Is this remote workstation? no
• How did you install Open3D? pip

Additional context
See linked issues.

\cc @sanskar107 @yxlao