Make torch an optional dependency

[demetd]

Is your feature request related to a problem? Please describe.

The torch package, which is currently a required dependency for causalml, requires presence of CUDA libraries like libcublas.so.*[0-9]. This causes installation issues, especially in environments where these libraries are not pre-installed or compatible, leading to errors such as ValueError: libcublas.so.*[0-9] not found in the system path. However, within causalml, torch is only used for the Causal Effect VAE model, which relies on neural networks.

torch is also a large package, would be good to not have to install it by default.

Describe the solution you'd like
To resolve this, I propose making torch an optional dependency within causalml via pip install causalml[torch]. This would be similar to the previously implemented approach with TensorFlow in #343, where it was made optional and only installed when needed via pip install causalml[tf].

Describe alternatives you've considered

• A workaround for envs with no GPUs is to specify the following in your Pipfile when you want to add causalml:

[[source]]
name = "pytorch"
url = "https://download.pytorch.org/whl/cpu"
verify_ssl = true

torch = {version = "==2.1.2", index = "pytorch"}

• A better, but way more involved, alternative would be to separate training deps from prediction deps as suggested in: Reduce number of 3rd party packages required for a prediction-only setup  #594

[demetd]

Thank you for the great and super fast work on this, @jeongyoonlee! Do you have an ETA for causalml's next release?