Add configurable CorrDiff regression loss selection

[younes-abid]

PhysicsNeMo Pull Request

Description

This PR adds configurable regression loss selection for the CorrDiff training
path when using regression and lt_aware_regression model names.

What changed

• Added a configurable loss field in CorrDiff regression model configs.
• Added loss selector wiring in examples/weather/corrdiff/train.py, with
  RegressionLoss as default to preserve existing behavior.
• Added robust selectable regression loss variants:
  • RegressionHuberLoss
  • RegressionCharbonnierLoss
  • RegressionEdgeAwareLoss
  • RegressionShiftTolerantLoss
  • RegressionHybridStructuralLoss
• Exported the new loss classes through the diffusion metrics public API.
• Added tests covering the new selectable regression losses.
• Added a changelog entry for this user-visible feature.

Why

This improves flexibility for CorrDiff regression training by allowing loss
function selection from config without changing code, while preserving backward
compatibility through the existing default loss.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.
• An issue is linked to this pull request.
• If I am implementing a new model or modifying any existing model, I have followed the Models Implementation Coding Standards.

Dependencies

No new runtime dependencies were added.

Review Process

I understand this PR will be reviewed by the PhysicsNeMo team and that
AI-generated comments are advisory.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

CODEOWNERS review map

Current for commit 27606bef7387. An approval covers every file listed for that owner; one owner is sufficient for shared files.

⏳ @CharlelieLrt — 5 file(s)

• examples/weather/corrdiff/conf/base/model/lt_aware_regression.yaml
• examples/weather/corrdiff/conf/base/model/regression.yaml
• examples/weather/corrdiff/train.py
• physicsnemo/diffusion/metrics/init.py
• physicsnemo/diffusion/metrics/legacy_losses.py

No CODEOWNER

• CHANGELOG.md
• test/metrics/diffusion/test_losses.py

Comment /codeowners-info to refresh.

[greptile-apps]

Greptile Summary

Adds configurable CorrDiff regression-loss selection while preserving RegressionLoss as the default.

• Introduces Huber, Charbonnier, edge-aware, shift-tolerant, and hybrid structural regression losses.
• Exports the new losses through the diffusion metrics API and wires them into CorrDiff training configuration.
• Adds basic invocation and output-shape coverage for each selectable loss.
• Documents the selectable variants in the model configurations and changelog.

Important Files Changed

Filename	Overview
physicsnemo/diffusion/metrics/legacy_losses.py	Adds shared regression forwarding helpers and five robust or structural losses; the shift-tolerant variant performs independent per-pixel shift selection rather than coherent alignment.
examples/weather/corrdiff/train.py	Resolves configured regression loss names through a fixed allowlist and retains the original loss as the backward-compatible default.
test/metrics/diffusion/test_losses.py	Adds basic callability, tensor-type, and output-shape coverage, but does not distinguish coherent shift alignment from per-pixel matching.
physicsnemo/diffusion/metrics/init.py	Exposes the new regression losses through the existing diffusion metrics public API.
examples/weather/corrdiff/conf/base/model/regression.yaml	Adds the default regression loss setting and documents available selectors.
examples/weather/corrdiff/conf/base/model/lt_aware_regression.yaml	Adds the same backward-compatible loss selection to lead-time-aware regression.

_{Reviews (1): Last reviewed commit: "Add configurable CorrDiff regression los..." | Re-trigger Greptile}