Translate a lone TTY ESC to the escape key in terminal-input buffers - #595

Problem

On a TTY frame, pressing ESC in a ghostel buffer sent nothing: the mode maps bind <escape>, a function-key event legacy TTYs never generate, while the TTY delivers the raw byte 27. The M-<char> bindings make [27] a prefix keymap, and Emacs has no ESC timeout, so read-key-sequence blocked with [27] pending forever. Collateral: ESC ESC also wedged (global ESC ESC ESC prefix), and C-g pressed while ESC was pending completed as C-M-g, injecting ^[^G into the PTY.

Fix

• New ghostel-tty-escape-delay (default 0.01 s, nil disables): a terminal-local menu-item :filter on [?\e] in input-decode-map (the evil-esc-mode technique) yields [escape] when no follow-up byte arrives within the delay — only in ghostel terminal-input buffers (semi-char/char with a live term). Everywhere else the filter returns the saved map, so escape sequences, ESC-as-meta, and non-ghostel buffers decode byte-identically to before. The translated event goes through the native key encoder, so kitty-protocol children get CSI 27 u and legacy children a raw ESC.
• Installer runs from the ghostel-mode body and from a buffer-local window-buffer-change-functions handler (covers later emacsclient -t frames). TTY terminals only; idempotent per terminal; re-wraps if another package replaces the entry. It reads the raw entry structurally because lookup-key resolves menu-item filters (which would drop e.g. evil's wrapper), and detects its own wrapper by the :filter tag because define-key copies menu-item lists.
• ESC ESC is bound to the key encoder as alt+escape: the filter deliberately leaves the second ESC of a fast pair raw, so the pair reaches [27 27] instead of pending on the global keyboard-escape-quit prefix.
• Composes with evil's evil-esc filter in either install order; at most one translation delay is paid per key.

Verification

• 8 new untagged ERT tests (filter guard matrix, installer wrap/idempotence/re-wrap/GUI-skip, mode-body integration pin, ESC ESC bindings); existing ESC decode invariants (C-[ not hijacked, ESC [/ESC O unbound) still pass. make -j8 all green.
• Live-verified in sandboxed TTY Emacs 31: lone ESC → ^[ promptly in semi-char and char with no pending prefix; C-g clean; fast ESC x → ^[x; arrows → ^[[A; slow ESC ESC → two ^[; fast ESC ESC → ^[^[; kitty protocol → ^[[27u; non-ghostel buffers and M-x untouched. GUI <escape> regression-checked.
• evil-ghostel live matrix: filters compose (ghostel wrapping evil's entry), evil's ESC still works in plain buffers, primary-screen ESC goes to evil, alt-screen ESC goes to the TUI, fast C-c M-d still decodes as meta, C-c + paused ESC runs evil-force-normal-state.

Known limitation (deliberately deferred): no defining-kbd-macro fixup à la evil — a macro recording a lone ESC records the raw byte and may not replay as escape.