Fix stale LLM node running state

[MEIS-DaCaiTou]

Summary

Fixes stale "running" indicators on Smart Canvas LLM prompt nodes.

When an LLM prompt node starts, the UI sets node.running = true. If the canvas is saved while that request is in flight, the transient running state can be persisted into the canvas JSON. The LLM completion path clears the in-memory flag, but did not schedule another save after node.running = false, so a refresh or later reopen could continue to show nodes as running even when no request is active.

This also filters chat model names containing * from the model selector. Values like gemini-3-pro-preview-thinking-* behave like wildcard placeholders rather than concrete model IDs, and can lead to provider errors when selected as-is.

Changes

• Clear persisted running flags before saving canvas nodes.
• Clear any stale running flags while loading a canvas.
• Schedule a save after temporary running indicators are cleared.
• Filter non-concrete chat model placeholders containing * from chat model options.

Validation

• Ran node --check static/js/smart-canvas.js successfully.
• Reproduced the symptom locally as Smart Canvas LLM nodes appearing to run indefinitely despite no active upstream request.