Improve documentation for loading pretrained weights with timm universal (tu-) encoders

soft_tversky_score doesn't use abs

Inconsistent `from_logits` parameter across loss functions

tu-convnextv2_base cannot be combined with UnetPlusPlus

VIT Adapter

segformer model implementation != original arch design

Possibility of optimisation in Focal Loss implementation

Example "Train multiclass segmentation on CamVid" Image normalization may be error

Segformer output not determinsitic

ValueError: invalid type: 'torch.mps.FloatTensor'