Fix conversion of ComfyUI v3 (io.ComfyNode) nodes

[fanfu70]

The planner only understood the legacy V1 node API and mis-converted v3 (comfy_api.io.ComfyNode) nodes in two ways:

1. Dynamic inputs (e.g. the DynamicCombo on ResizeImageMaskNode) are sent by the frontend as dotted prompt keys ("resize_type" plus "resize_type.shorter_size"). The planner filtered inputs against the execute() signature, but v3 nodes expose EXECUTE_NORMALIZED(*args, **kwargs), so the filter was silently disabled and the dotted keys were emitted as literal kwargs. Generated scripts crashed with: TypeError: ResizeImageMaskNode.execute() got an unexpected keyword argument 'resize_type.shorter_size'

2. Hidden inputs (UNIQUE_ID / PROMPT / EXTRA_PNGINFO) are not execute() kwargs for v3 nodes; the executor injects them on a per-node class clone (PREPARE_CLASS_CLONE -> cls.hidden). The planner instead injected fabricated kwargs such as unique_id=, which v3 execute() rejects, and left cls.hidden as None, crashing nodes that read hidden inputs (e.g. GetImageSize's cls.hidden.unique_id) in headless scripts.

Changes in comfyui_to_python/generator/planner.py:

• Detect v3 nodes via FUNCTION ("EXECUTE_NORMALIZED" / "EXECUTE_NORMALIZED_ASYNC") or the presence of EXECUTE_NORMALIZED. Async v3 nodes raise a clear error: generated scripts are synchronous.
• prepare_v3_node_call(): mirror comfy_execution.get_input_data() using ComfyUI's own machinery (comfy_api.latest._io):
  • get_finalized_class_inputs() expands dynamic schema entries against the live prompt keys,
  • build_nested_inputs() folds the dotted keys into the single nested-dict kwarg execute() actually receives, e.g. resize_type={"resize_type": "scale shorter dimension", "shorter_size": 1080}
  • stale dotted keys from unselected DynamicCombo options are dropped, and ACCEPT_ALL_INPUTS nodes pass extra inputs through as the executor does.
• Render v3 calls as type(obj).PREPARE_CLASS_CLONE({...}).EXECUTE_NORMALIZED(...) so cls.hidden is populated exactly like in the server executor. The payload carries "PROMPT" / "EXTRA_PNGINFO" when declared; UNIQUE_ID and DYNPROMPT are intentionally omitted because a headless script has no PromptServer (HiddenHolder resolves them to None).
• format_structured_value(): render prompt-derived list/dict literals (folded dynamic inputs, DICT widget values) as Python source, including link placeholders resolved inside nested dicts.

V1 nodes are untouched: signature filtering and unique_id / prompt / extra_pnginfo kwarg injection behave exactly as before (verified for **kwargs nodes such as VHS_LoadVideo / VHS_VideoCombine).

tests/runtime/generated/upscale-model-loader.py is regenerated with the fixed renderer (v3 clone-style calls; also picks up pre-existing drift in the inlined cleanup_comfyui_runtime helper from an older revision).

Verified:

• Full existing unit suite passes (codegen regression, upscale export, project contracts, node runtime cleanup, runtime validation harness).
• Regenerated a real LTXV video-enhancement workflow containing v3 nodes (ResizeImageMaskNode x2, GetImageSize and others): nested dicts and clone calls are correct, V1 rendering unchanged, output compiles.
• Executed the generated v3 call lines verbatim against the real node classes with synthetic tensors: correct output shapes (short side == 1080; multiples of 32) and GetImageSize runs headless without a PromptServer; the old dotted-kwarg form still raises the original TypeError.