feat: quick generation settings panel in chat

[sakebomb]

Summary

• Adds a temperature pill (T: X.X) in the chat input control bar showing the current session temperature (shown when a session is active and a model is loaded).
• Tapping the pill opens a compact bottom sheet with sliders for temperature and top-p, and an Unlimited toggle + max-tokens slider.
• Changes persist to the session's completion settings immediately via updateSessionCompletionSettings (which flips settingsSource to custom).
• Reset to defaults restores the system defaults.
• No new native dependencies.

Max tokens / unlimited

n_predict defaults to -1 (unlimited) since #687 (with a migration flipping existing 1024 → -1). A 64–8192 slider can't represent unlimited, so max tokens is an Unlimited toggle plus a finite slider — mirroring the full Generation Settings sheet's semantics. Turning the toggle off starts the slider at the same 1024 fallback the full sheet uses. This avoids the panel showing a false "64" cap on a default (unlimited) session and silently capping generation on any slider touch.

Test plan

• Temperature pill appears when a session is active and a model is loaded; hidden otherwise.
• Tapping the pill opens the QuickGenSettingsSheet.
• Sliders load the current session's completion settings.
• Apply saves settings and closes the sheet.
• Unlimited session → toggle on, slider hidden; toggle off → slider appears at 1024; apply preserves -1 when left on, caps when turned off.
• Reset restores the unlimited default (temp 0.7, top-p 0.95, max tokens unlimited).
• No active session → falls back to defaults.
• tsc, eslint, l10n:validate clean; full suite 269 suites / 4165 passed.
• Manual device pass (iOS/Android).

Closes #605

[sakebomb]

Rebased onto current main and did a self-review pass. One real defect surfaced that only appears against current main:

n_predict now defaults to -1 (unlimited) (via #687 + a migration flipping existing 1024 → -1). The original max-tokens slider was 64–8192, so on a default session it displayed a false "64" cap, and any drag would silently cap an otherwise-unlimited session. Fixed by making max tokens an Unlimited toggle + finite slider, mirroring the full Generation Settings sheet (finite fallback 1024).

Also corrected here: the shared test mock's defaultCompletionSettings.n_predict was stale at 1024 (now -1, matching the real default), and the sheet's tests were rewritten to actually exercise the unlimited/finite/toggle-off/reset paths (the old ones asserted a 1024 default that no longer exists). Both the toggle-save and slider-visibility guards are mutation-verified.

Verification: tsc clean · eslint clean · l10n:validate valid · full suite 269 suites / 4165 passed. No native changes.