Rebuild an open scene tab when a writer rewrites its scene

[saworbit]

Fixes #1068

Problem

project_rename_references and project_apply_changes rewrite .tscn files. The refresh from #1047 reloads what the editor's resource cache holds, and an open tab is not in it. The tab is a node tree the editor built from the file, so it kept the old scene, and the next editor_save_scene wrote that back over the file. The writer said success and editor_copies_reloaded: [].

What changed

• The bridge's resource.refreshCached now also rebuilds the tab of a path open in the editor, with EditorInterface.reload_scene_from_path, and says so per path. It checks the reload took by the tab's root: a reload that happened replaced it.
• A new bridge read, editor.openScenes, answers the open tabs and the unsaved report editor.getState already carries.
• Before writing, both writers ask it. A rewritten scene open with unsaved changes refuses with 409 unsaved_changes. Before Godot 4.7 the engine cannot say, so any open one refuses with 409 dirty_state_unavailable. Both name the scenes and carry discard_unsaved: true in retry_with. Nothing is written. The confirmation previews make the same check, so they refuse what the call would.
• discard_unsaved is a new boolean on both tools. It is the caller accepting the loss of the tab's unsaved changes, the same word scene_close uses.
• Answers gain editor_scenes_reloaded. A tab left stale is named in editor_copy_errors with the reason and the way out.
• viewport_create_test_lab rebuilds an open lab tab on overwrite, the way scene_create already reloads the scene it overwrites.
• Along the way: the rename's refusal of a new_name already in use named no fix. The new harness block was the first to reach it, and the remedy check failed on it. It now names new_name as the field.
• The discard_unsaved description is one short sentence. The first draft put the full listing 380 bytes over the tools/list budget, and the refusal already explains the argument where an agent reads it.

Engine facts this rests on

Measured with a throwaway editor plugin on 4.5.1, 4.6.2 and 4.7.2:

• reload_scene_from_path is synchronous. It returns with a new root holding the new file, keeps whichever tab was current, and drops the tab's unsaved changes and undo history. On 4.7 the scene leaves get_unsaved_scenes.
• On 4.5 and 4.6 a second reload in the same frame does nothing and prints nothing. So the bridge rebuilds one tab a request and answers the rest pending, and the server asks again. Commands are taken off the queue before any runs, so each request lands on a later frame.
• On 4.5 and 4.6 a reload of a path no tab holds clears the current scene's undo history, with no message. 4.7 prints an error instead. So the bridge only asks for paths get_open_scenes lists.

Tests

• Native: a rename rebuilds two clean open tabs across the pending retry. A rename refuses an unsaved tab and an unknown one, writes nothing, and goes through with discard_unsaved. Apply refuses the same way before its check runs. A tab that could not be rebuilt is named. The check asks only when it can matter.
• Live harness, tests/scene_tab_reload.ps1: the steps from the issue. A scene with a connection is opened and edited, the rename refuses, the scene is saved, and the rename runs (with discard_unsaved before 4.7). Then an edit and a save. The file keeps the new method name and the edit, on all three lines.