Include Max and Ultra in the inline reasoning-effort list, configurable, instead of only via 'More reasoning…'

[TUI] Include Max and Ultra in the inline reasoning-effort list (optionally, via config)

Environment: CLI 0.147.0, Windows, npm install.

Observed: In the TUI reasoning-effort popup (/model → model → reasoning level), the list shows only standard efforts (low/medium/high/xhigh) plus a single "More reasoning…" item that opens a second, advanced popup containing Max and Ultra. The Alt+, / Alt+. shortcuts similarly stop at Extra high and never step into Max/Ultra (per tui/src/chatwidget/reasoning_shortcuts.rs: "Raising never silently crosses into Max or Ultra; those efforts require the explicit advanced-reasoning picker").

I understand the intent (avoid accidentally selecting the most expensive efforts), but the indirection is confusing on models whose ladder genuinely tops out at Max (e.g. GPT-5.6-luna: low/medium/high/xhigh/max), where the "inline list max" is actually the model's real top rung.

Request: make Max/Ultra accessibility configurable rather than hardcoded:

• A config flag (e.g. tui.include_advanced_reasoning_efforts_inline = true, default false) that renders Max/Ultra inline in the standard popup (with a rate-limit warning attached), and
• optionally, a shortcut variant (or a pref/setting on Alt+,/Alt+.) that continues past xhigh into Max/Ultra with the existing confirm-style guard kept (e.g. a one-key confirmation or keeping the popup-only path when the flag is off).

Related: #32874 (Shift+Down no longer cycles to Ultra), #35292 (Ultra reset to xhigh on edit/resume).

This is a TUI-only request; the desktop app already exposes these efforts under its own enabled-list settings.

Start with tui/src/chatwidget/reasoning_shortcuts.rs and trace the TUI reasoning-effort popup and its configuration handling. Check how the standard and advanced lists are built, then define the configurable inline behavior and shortcut guard described in the issue. Done means Max and Ultra remain hidden by default, can be enabled inline through configuration, and shortcut behavior preserves confirmation safeguards.