LayeredWeightedPhysicsLoss reads layer_residuals off the model instead of receiving it

[albanpuech]

Low priority / cleanup. Pre-existing design issue, no known incorrect results today. Surfaced while reviewing #87.

GNS_heterogeneous.forward writes self.layer_residuals as a side effect, and LayeredWeightedPhysicsLoss (training/loss.py:363) reads model.layer_residuals afterwards. The data dependency is invisible in both signatures — the loss just receives model=self.model.

Concrete downsides:

• Unenforced ordering. Requires forward() before loss_fn() on the same instance. If the dict is empty, L = 0 produces no loop iterations, total_loss stays a Python float, and it fails on total_loss.item() — an obscure error far from the cause.
• Not reuse-safe. One dict per module instance: two forward passes before a loss call silently overwrite, so gradient accumulation or evaluating two batches before reducing would read the wrong batch's residuals and return a plausible number rather than erroring.
• Hidden model/loss coupling. Only works with models exposing layer_residuals, i.e. GNS_heterogeneous; AttributeError on GRIT. A config pairing GRIT with this loss fails at runtime, not at validation.

Working as intended today: the stored tensors keep their autograd graph (no .detach()), so gradients flow correctly.

Suggested fix: return residuals from forward alongside the predictions (as #87 now does for embeddings) and pass them into the loss explicitly, so the dependency is in the signature and the model= kwarg is no longer needed for this.

Deferred from #87 because the fix touches the shared BaseLoss.forward signature across all losses.

Refs #87