Add GPU support for libtorch within MOOSE #26718
Description
Reason
MOOSE has been relying on CPUs, but GPUs can considerably accelerate libtorch-based neural network evaluations on both Mac and Linux distributions. This would reduce expensive data-driven material model evaluations.
Design
The main points:
- Transitioning to conda-based packages on Macs and adding MPS support on ARMs
- Creating a CUDA environment that can be used both through conda and in a container for Linux boxes
- Option to select the used device within the App combined with automatic device-detection routines.
Impact
Will accelerate the training and evaluation processes of libtorch-base neural networks.