Variant C: Add on_blur_by_user and migrate blur-cancelling controls onto it

[butvinm]

Review-only PR in the fork. Do not merge upstream from here. One of three competing designs for the same problem, built so they can be compared side by side. See also the variant A and variant B PRs.

Base is fix-layout-switch-cancels-rename (merged upstream as zed-industries#61852), so the diff shows only this design.

Design: leave GPUI alone, make the correct pattern the easy one

Two subscription constructors on Context<T>, placed immediately beside the two methods that carry the trap (crates/gpui/src/app/context.rs):

pub fn on_blur_by_user(&mut self, handle: &FocusHandle, window: &mut Window,
    listener: impl FnMut(&mut T, &mut Window, &mut Context<T>) + 'static) -> Subscription

pub fn on_focus_out_by_user(&mut self, handle: &FocusHandle, window: &mut Window,
    listener: impl FnMut(&mut T, FocusOutEvent, &mut Window, &mut Context<T>) + 'static) -> Subscription

Each is a three-line wrapper gating the existing constructor on window.is_window_active(). GPUI's event semantics are entirely untouched, so this is the lowest-risk option to land.

Migrated: project panel, collab panel, Go to Line, picker (both head kinds), terminal tab rename, and modal_layer. ErasedEditorEvent::Blurred was deleted once it had no consumer left - the one place the convention became an actual constraint.

Diffstat

13 files changed, 368 insertions, 65 deletions.

Verification

• Implementation branch: https://github.com/butvinm/zed/actions/runs/30459148116 - cargo check -p zed --all-targets passing; the test step failed, but only because this branch's own test repair was missing workspace.add_panel(...), so its precondition assertion fired.
• Against the corrected deactivation test (branch focus-variant-c-cmp): https://github.com/butvinm/zed/actions/runs/30462742452 - passing, test_rename_survives_window_deactivation ... ok. The implementation is fine; the earlier red was a test-harness artifact.

Findings worth more than the code

• modal_layer is not the good prior art it appears to be. It has no activation awareness and only looks correct because dismiss_on_focus_lost is usually false. file_finder's on_before_dismiss does set it, so that was a real latent bug, fixed here.
• The base branch's regression test was vacuous - this is what surfaced it. Fixed properly in the separate fix-vacuous-deactivation-test PR.

Honest weaknesses

• This is a convention, not a constraint. cx.on_blur, cx.on_focus_out, and EditorEvent::Blurred all still exist and still misbehave; 12 other Blurred subscribers are untouched. Nothing stops the next contributor from hand-rolling the buggy version. It does not close the bug class for future UI and cannot without removing or renaming the dangerous API outright.
• Callback ordering changed in the five editor migrations: they moved from EditorEvent::Blurred subscribers (which run in the effect flush) to focus listeners (synchronous inside Window::draw). For the picker, cancel/DismissEvent now runs before ModalLayer's focus-out listener rather than after. Wants hands-on exercise of the command palette, file finder, and file-finder submenu.
• Two new GPUI tests were added whose scaffolding (two track_focus divs producing the expected focus paths) is the kind of thing that is easy to get wrong on paper.

Full API rationale, migration-by-migration notes on how cleanly each call site fit, and all judgment calls are in VARIANT-NOTES.md at the repo root of this branch.

Release Notes:

• N/A

[butvinm]

Superseded by #5 (variant D). Closing to keep the field clear; the diff stays viewable here.