#2258 Add opt-in browser model downloads and switching through official setup

Descrição

Users currently run setup and separate start scripts to download or change models. This adds an opt-in Models table that stays available during installation and model loading, with download review, progress, start/stop, and switching between installed models.

Launch with `MODELS.bat` or `sh models.sh`. The Models page listens on `127.0.0.1:8090`; the normal chat and API server uses port 8080.

- Read regular model choices directly from `setup.py` and run `setup.py --setup --yes --no-start`. Leave GPU, backend, context, KV, low-RAM mode and cache settings to official setup. Pin the reviewed data folder and require an explicit download confirmation.
- Keep an existing shared Qwen lookup-table location after setup, checking shard metadata and backing up the generated config before changing paths. Swift and Unsloth keep their own tables. OrcaRouter and experimental models link to their documented manual workflows.
- Reuse the existing management helper's process tracking, start/stop and resumable setup jobs. Refuse to interrupt busy models or stop servers started elsewhere; check target files before switching and attempt to restore the previous model if loading fails.
- Keep management local, validate model ids and reviewed plans, and require same-origin JSON requests plus a per-launch token. Recover pending setup on reopening the launcher and allow only one Models launcher per checkout.

The normal installer, default web app, server and engine backends are unchanged. Their files have no diff against v0.1.43.1. This is one optional launcher feature; it does not change downloader behavior, add inference tuning, or automate OrcaRouter's conversion.

Validation: 48 tests passed across the new launcher and existing MCP management suites, using fake setup/server processes. Existing MCP tests' ambient port-8080 probes were isolated from an already-running local server. Python compilation, JavaScript syntax and diff checks passed. Edge rendering at 1440px and 390px had no JavaScript errors or page overflow; download review, model labels and keyboard focus across polling were checked. No real model download or GPU inference was performed; no C++/CUDA/HIP/SYCL files are changed.

Related to #925 (browser feedback during setup). Usage and limitations are documented in `docs/MODEL_MANAGER.md`.