Update PyTorch 2.2.0, Scilpy 2.0.2, DWI_ML and adapt target SH order based on the training data.

[levje]

PyTorch update to 2.2.0

Didn't change anything except the install.sh file to download another PyTorch version and support different CUDA versions (w.r.t. PyTorch docs). Seems to works fine.

Scilpy 2.0.2

A few changes here and there to adapt with functions that were moved/removed from scilpy. Logic shouldn't have changed, I used the new functions as they are used in some scilpy scenarios/tests. Changing to scilpy 2.0.2 is useful since its wheel is available on Compute Canada.

DWI_ML

IMPORTANT: Waiting for scil-vital/dwi_ml#234 before merging. Might have to adapt the commit hash in the requirements.txt to refer to the main once that PR is merged.

Update of DWI_ML is useful because it now uses PyTorch 2.2.0 and Scilpy 2.0.2 which is needed for Compute Canada.

SH Order changes:

During training:

• When loading the first subject to initialize the environment, use the input_volume from that subject to extract sh_order that the model will be trained on.
• Save the sh_target_order to the hyperparameters.json file.
• I adapted the default example model's hyperparameters.json file accordingly.

During tracking/testing (with ttl_track*):

• Load the sh_target_order directly from the hyperparameters.json file. This way the networks are initialized with the input size corresponding to the one from the pretrained model.

Other:

• Also removed self.interface_seeding = hyperparams['interface_seeding'] and self.add_neighborhood = hyperparams['add_neighborhood'] since they didn't seem to be used anymore and they caused issues when training/tracking.

[AntoineTheb]

Looks good ! I'll do a proper review once I'm back. Ideally, we would set the option added in this PR: scil-vital/dwi_ml#238 to False as I realised when profiling that cache clearing slows down tracking a lot. I'll try to finish this PR asap.

[levje]

Checks will fail, but once the clearing cache PR in DWI-ML is merged, we should be good to merge that one if everything is ok. Also, make sure that the clearing_cache option is merged into the for_beluga_scilpy2 branch.