Recipes: Add "Train on this dataset" card with training configuration

[LeoBorcherding]

Closes #6178

Summary

Adds a first-class Train node to the Recipe Studio graph. A finished dataset pipeline connects into this card, which holds a full training config and kicks off a fine-tuning run on the generated output. It's the first step toward the closed loop: generate, train, benchmark, regenerate.

What's included

• New train node kind with its own card, built on the existing BaseNode primitives so it matches the other nodes (fuchsia tone, stat chips, dataset-status row, primary action button). A single data-input handle lets the recipe output wire straight in.
• Dataset auto-wiring. The card defaults to "Auto-wire from recipe output": it reads the most recent completed execution's artifact_path (RecipeExecutionRecord) and trains on that. You can override it with a Hugging Face dataset or a local file.
• Full training config in the Configure dialog, mirroring the Studio training section. Base model, method (QLoRA/LoRA/Full/CPT), epochs, LoRA rank, batch size, learning rate, max sequence length, and output adapter name sit up front. Everything else (LoRA alpha/dropout/variant, gradient accumulation, warmup, max steps, weight decay, optimizer, scheduler, gradient checkpointing, seed, packing, train-on-completions, W&B/TensorBoard, HF token) lives behind an Advanced accordion, the same pattern the other cards use.
• Inline launch. "Run training" maps the card config into the shared training-config store and calls the existing startTrainingRun(), so it reuses the real training runtime, consent gates, and dataset checks. The button disables while a run is active and surfaces start errors on the card.
• Palette, drag-and-drop, and round-trip. Train shows up in the Add-a-step sheet, supports drag-drop, and survives recipe JSON export/import. It's persisted UI-side like markdown_note, so it stays out of the dataset payload but is restored on load.

Data wired

• Recipe output: RecipePayload.run.artifact_path, run.dataset_name, run.output_formats
• Training input: TrainingConfigState.datasetSource, uploadedFile, dataset, datasetFormat

Testing

• npm run typecheck: clean
• biome check: new files clean, no new lint introduced
• Builds and runs from source in a local studio (supervisor install.ps1 --local). Screenshots to follow.

The Train node is UI-only in the recipe payload (not a dataset column); the backend training API is unchanged.