menu: Focus a context menu before its first frame

[madcodelife]

Closes #3364

Description

With accessibility active, opening a ContextMenu aborts a debug build with set_focus called more than once in a single frame.

GPUI registers the focused element as the frame's accessibility focus while prepainting it, and allows one per frame. ContextMenu moved focus to its menu inside the menu's own prepaint. The menu is deferred and prepaints last, so the element that held focus before (for example the right-clicked input) had already registered, and the menu registered a second time.

• ContextMenu now focuses the menu in the deferred callback that builds it, before its first frame is drawn, after the input selection ownership from input: Preserve selection, focus ring and context menu actions #3382 is resolved. DeferredMenu::build_menu no longer touches focus.
• The old prepaint focus ran on every frame and pulled focus back whenever it left the menu. That also covered a case it was never written for: entering a submenu with the keyboard and then pointing at another parent item closes the submenu while it still holds focus, leaving Escape and the arrow keys with nowhere to go. PopupMenu now takes focus back itself when a hover closes a focused submenu, which also applies to DropdownMenu and AppMenuBar.

Behavior change: while a context menu is open, focus that the application moves elsewhere is no longer pulled back to the menu on the next frame.

How to Test

cargo test -p gpui-component --lib menu::
cargo test -p gpui-kit --features test-support --test menu

• menu_takes_focus_before_its_first_frame checks that focus has left the previous element when the frame that draws the menu begins rendering. It fails on main.
• hovering_away_from_a_keyboard_entered_submenu_refocuses_its_parent enters a submenu with the keyboard, hovers another parent item, and checks that Escape still dismisses the menu. It passes on main, fails with only the ContextMenu change, and passes with this PR.

The headless test platform cannot activate the accessibility tree, so the tests cover the focus ordering that causes the panic rather than the panic itself. This has not been run against a screen reader on a native debug build.

Checklist

• I have read the CONTRIBUTING document and followed the guidelines.
• Reviewed the changes in this PR and confirmed AI generated code (If any) is accurate.
• Passed cargo run for story tests related to the changes.
• Tested macOS, Windows and Linux platforms performance (if the change is platform-specific)