Add VolumeConservationLoss to physicsnemo.experimental.losses

[wdyab]

Summary

Adds VolumeConservationLoss to physicsnemo.experimental.losses: a generic,
dimension-agnostic weak-conservation loss. It compares the (optionally
volume-weighted) spatial integral of pred and target at each step of a
trailing feature/time axis, using a dependency-injected data-fitting metric
callable and an optional inactive-cell mask.

The default metric is mse (always available); any physicsnemo.metrics.general
function — or a relative-error metric such as the one proposed in #1746 once it
lands — can be dependency-injected. Inputs are validated (pred/target shape,
cell_volumes shape, and static/per-sample mask shape) under a
torch.compiler.is_compiling() guard; the masking path uses static-shape
weighting (no boolean indexing), so it is torch.compile-compatible.

Named volume conservation (not "mass conservation") on purpose: it conserves
a volume-weighted spatial integral, not true mass — no material density or
porosity is applied.

Stacked PR. Depends on #1808, which creates the physicsnemo.experimental.losses
package. This branch is based on nof/pr3b-spatial-derivative-loss, so the diff
also shows #1808's files until it merges; the change owned by this PR is the top
commit (physicsnemo/experimental/losses/volume_conservation.py + its test, plus
the __init__ export). Once #1808 merges I'll rebase this onto main.

Closes #1806

Test plan

• test/experimental/losses/test_volume_conservation.py: analytic 2D/3D
  reservoir integrals (known totals), dependency-injected metric sweep,
  volume-weighted vs uniform, multi-timestep, masking (static + per-sample),
  gradient flow, scale invariance, uniform-default, and error paths
  (pred/target, cell_volumes, and mask shape mismatches).
• interrogate 100%, ruff check+format clean, doctest, license headers, CHANGELOG.

[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.

[greptile-apps]

Greptile Summary

This PR adds experimental weak-physics losses for conservation and derivative regularization. The main changes are:

• New physicsnemo.experimental.losses package exports.
• Added VolumeConservationLoss for volume-weighted spatial integrals.
• Added SpatialDerivativeLoss and central-difference helpers.
• Added tests for metrics, masking, gradients, and validation paths.

Important Files Changed

Filename	Overview
physicsnemo/experimental/losses/volume_conservation.py	Adds VolumeConservationLoss with shape validation, optional volume weights, and per-sample or static masking.
physicsnemo/experimental/losses/derivative.py	Adds SpatialDerivativeLoss with central-difference helpers and mask-aware stencil selection.
physicsnemo/experimental/losses/init.py	Exports the new experimental loss classes and derivative helpers.
test/experimental/losses/test_volume_conservation.py	Adds coverage for volume conservation behavior, masking, custom metrics, gradients, and shape errors.
test/experimental/losses/test_derivative_loss.py	Adds coverage for finite differences, derivative loss behavior, masking, gradients, and shape errors.

_{Reviews (2): Last reviewed commit: "Add VolumeConservationLoss to physicsnem..." | Re-trigger Greptile}

[wdyab]

@peterdsharpe @mnabian @melo-gonzo for your review

[wdyab]

Closing per the offline discussion with the maintainers: these regularization losses will live in the reservoir-simulation example recipe downstream, and the generic finite-difference need will be addressed by extending physicsnemo.nn.functional.derivatives (non-periodic + batched) rather than adding a parallel implementation. Thanks for the reviews!