fix: show cold-start loading label during engine warmup

[quiet-node]

Description

While the built-in llama.cpp engine or a local Ollama model is cold-starting, the chat view showed only the bare typing dots with no indication of what was happening. This adds a phased loading label next to the dots that narrates the wait, and marks the engine warmed the instant a real chat request's first token proves it (rather than waiting on a separately-queued warm-up prime).

Key changes

• useEngineLoadingLabel hook: shows nothing for sub-second waits, then cycles phase-1 filler ("Starting up the model…", "Reading model weights…"), and jumps to phase-2 filler ("Warming up…", "Bigger models take a little longer…") the instant the built-in engine's real warmup:builtin-warming signal fires. Progress is monotonic — once phase 2 is entered it never falls back to phase 1.
• useEngineWarmupStatus hook: app-root-mounted subscriber to the existing warmup:builtin-warming/-warmed events, shared with the Settings status line.
• ConversationView: wires the label into the existing loading row, deferring to an active search-stage label when one is present. Remote (openai-kind) providers never show a label.
• Backend: stream_builtin_chat now marks the engine warmed off the real request's own first streamed token (BuiltinWarmState::mark_warmed_by_real_request), independent of a proactive prime that may still be queued behind it at the engine's single execution slot. This fixes the Settings "warming" status getting stuck for the duration of a response that raced ahead of its own prime.

Testing

• bun run test:all:coverage — frontend 100/100/100/100, backend --fail-under-lines 100, all passing.
• bun run validate-build — clean (lint, format, typecheck, build).