Improve performance of Tensorflow and PyTorch modules

[stotko]

Retrieving properties from Tensorflow and Pytorch in our find modules is expensive because calling Python and importing either of them comes with a significant cost. Improve the performance by fetching all required properties in a single pass instead of 1 property per pass:

time cmake -DCMAKE_INSTALL_PREFIX=../bin -DBUILD_CUDA_MODULE=ON -DBUILD_UNIT_TESTS=ON -DBUILD_BENCHMARKS=ON -DBUILD_PYTORCH_OPS=ON -DBUILD_TENSORFLOW_OPS=ON -DBUNDLE_OPEN3D_ML=ON -DOPEN3D_ML_ROOT=../../Open3D-ML -DOPEN3D_THIRD_PARTY_DOWNLOAD_DIR=~/workspace/open3d-downloads -DBUILD_JUPYTER_EXTENSION=ON -DBUILD_RPC_INTERFACE=ON -DBUILD_LIBREALSENSE=ON -DBUILD_AZURE_KINECT=ON -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ..

Before

real 0m19,039s
user 0m21,909s
sys 0m22,577s

After

real 0m9,561s
user 0m8,882s
sys 0m6,575s

In a future PR, the performance could be improved even further by only calling the find modules once instead of twice.

---

This change is 

[update-docs]

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.

[benjaminum]

Reviewed 2 of 2 files at r1.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @stotko)