Always leave read-only mode on exit; keep evil insert out of it - #716

Fixes #715.

Core

ghostel-readonly-exit always leaves read-only mode. Switching between copy and Emacs mode keeps the mode active before the first read-only entry, so Emacs → copy → exit lands in semi-char (or char), not back in Emacs mode. A key typed in copy mode entered from Emacs mode now reaches the terminal instead of being dropped by the fast-exit handler.

ghostel--cursor-row-input-start holds the cursor-row prompt detection (OSC 133 prop, then ghostel-prompt-regexp) previously duplicated in ghostel-input-start-point and line mode's startup check.

ghostel-prompt-navigation-input-mode accepts nil.

evil-ghostel

• Entering insert state from copy or Emacs mode exits to the terminal. Point's column on the cursor row is kept and clamped to the input start after a detected prompt; off the row, point lands on the cursor. Emacs state (C-z) keeps copy mode.
• In normal and motion state, ghostel-prompt-navigation-input-mode and ghostel-point-leave-input-mode are nil buffer-locally, so [[ / ]], / search, and minibuffer landings stay in semi-char, where the terminal-aware j / G / i / a work. Insert and Emacs state keep ghostel's default copy-mode switch.

Note on the issue's diagnosis: under emacs -Q the fast-exit remap still fires in insert state, so the "every key is read-only" symptom depends on the user's config (or ghostel-readonly-fast-exit nil). Insert state in a read-only buffer is wrong either way.

Testing

• make -j8 all passes; new ERT tests run in CI's module-less test-evil job and fail without the fix.
• New elate suites test/elate/matrix/readonly-evil-ghostel.json and readonly-ghostel.json (zsh) pass every group.
• Existing elate matrix (shells ×4, python3, boundary ×4, tracking, word-boundaries, scrollback, kitty) shows no regressions. ~ XPASSes on every shell on main as well (it no longer signals read-only, but the edit still does not land); P on nu XPASSes on this branch.