Replace step return tuple with a StepOutput dataclass carrying pre-correction values

[jpdunc23]

What to build

Replace the tuple[TensorDict, StepperState | None] return of the step abstraction with a StepOutput dataclass, and plumb pre-correction values up through the stepper so prediction BatchData optionally carries them. Pure plumbing: no user-visible behavior change.

This is PR 1 of a 3-PR split of #1218 and #1222, both of which predate the StepperState refactor on main. Branch from current main HEAD. Follow-up issues cover correction inference metrics and corrector training features.

Design decisions (settled in design review):

• Named StepOutput, following the repo's *Output convention (TrainOutput, LossOutput); *Result is reserved for benchmarks.
• Fields: output (denormalized post-adjustment data), stepper_state (unchanged passthrough semantics), and uncorrected.
• uncorrected carries pre-correction values (not deltas) of exactly the corrector-modified variables; empty dict when no corrector ran, so consumers need no None checks. Deltas are always derivable as output − uncorrected; values are not recoverable from deltas, and the training-features follow-up needs the values.
• StepperState is not used to carry uncorrected data: its contract is recurrent per-sample state threaded across predict windows, while uncorrected values are a per-window, per-timestep diagnostic series.
• Scope: the step package and corrector tap, the stepper's step/predict generator plumbing, and an optional uncorrected field on BatchData (following the data_mask/stepper_state optional-field precedent), populated only on prediction BatchData returned by predict. No return_uncorrected argument or overloads. PairedData exposes the uncorrected prediction for the metrics follow-up. If the BatchData layer bloats this PR, it may slip to the metrics PR.
• Uncorrected tensors are detached unconditionally here; attach/detach control arrives in the training-features follow-up with its first consumer.
• The output-process function (e.g. spatial masking) applies per-variable and name-preserving, so it is applied to the uncorrected subset identically.
• The rollout always feeds corrected output forward as state.
• fme.coupled touches only what is needed to unwrap StepOutput.output; all coupled feature support is deferred to a later PR.

Acceptance criteria

• Every step implementation returns StepOutput, with uncorrected populated exactly when a corrector modifies variables and an empty dict otherwise; stepper_state semantics unchanged. (Prior art: the step package's existing unit tests; updated versions exist in both source PRs.)
• predict returns prediction BatchData with uncorrected set; rollout state is the corrected output; uncorrected tensors are detached. (Prior art: existing stepper unit tests.)
• The existing distributed/spatial-parallel step test matrix passes (baselines generated single-rank, verified under torchrun, per repo convention).
• No user-visible behavior change; existing checkpoints load unchanged.

Blocked by

None - can start immediately.

Notes

Source PRs serve as reference implementations only; neither merges as-is (#1222 is explicitly marked don't-merge): #1222 for the corrector tap and step-return change, #1218 for its alternative StepResult framing. Where they disagree, the decisions above win.