Add IMPACT-Synth training config (losses + optimizer) so the app is fine-tunable

fix: log the image of the module the model exposes

Block-level UNetPlusPlus.yml + Tanh + clean-key checkpoints (forward-exact 0.0)

Use declarative UNetpp.yml + clean-key checkpoints (forward-exact 0.0, inference-only)

Use SAM_Perceptual(train=True, weights=[0,1,1,0]) as training loss instead of IMPACTReg

Update whole-body sCT model

Update CBCT/Config.yml

Update CBCT/Config.yml

Update CBCT/Config.yml

Update CBCT/Config.yml

Update CBCT/Config.yml

Update CBCT/Config.yml

Update CBCT/Config.yml

Upload Config.yml