Make loss weight masking conditional on loss function support

[donglaiw]

Summary

This PR refactors the deep supervision loss computation to conditionally apply per-voxel weight masks only to loss functions that explicitly support them, rather than unconditionally passing weight masks to all loss functions.

Key Changes

• Added _loss_supports_weight() utility function that uses introspection to check if a loss function's forward method accepts a weight parameter
• Updated three loss computation methods to conditionally apply foreground-weighted masks:
  • compute_multitask_loss()
  • compute_loss_for_scale()
  • compute_standard_loss()
• Loss functions that support weight masks (e.g., WeightedMSELoss, WeightedMAELoss, SmoothL1Loss) receive the mask
• Loss functions that don't support weight masks (e.g., MONAI DiceLoss, BCEWithLogitsLoss) are called without the weight argument

Implementation Details

• The _loss_supports_weight() function uses inspect.signature() to safely check the loss function's forward method signature
• Gracefully handles inspection errors by returning False as a fallback
• Foreground weighting (2.0x for SDT > 0) is only computed when needed, reducing unnecessary tensor operations
• This approach maintains backward compatibility while enabling use of diverse loss functions in the same training pipeline

https://claude.ai/code/session_01XRs3k4q3869VksCMeHsMTN