docs: Windows quick-start and a platform-aware doctor preflight

[luminary19]

Summary

Makes the doctor preflight platform-aware and adds a Windows quick-start to
the README.

scripts/doctor.sh:

• Python check tries python3, then python, then the py launcher (each
  verified to actually run), instead of python3 only.
• Detects Windows via uname -s (MINGW*/MSYS*/CYGWIN*) and, on Windows,
  prints the accurate GPU line — "CUDA llama.cpp serving available (vLLM/SGLang
  require Linux)" — and defaults the models dir to %USERPROFILE%\models
  instead of /models. The POSIX branch is byte-for-byte the previous behavior.

README.md:

• Notes Apple Silicon → MLX and Windows → llama.cpp/CUDA in the prereqs, and the
  %USERPROFILE%\models default.
• Adds a Windows subsection: Git-for-Windows/Bun/Node/Python prereqs, the
  PowerShell 5.1 && caveat, the services/node_modules junction (no Developer
  Mode needed), the llama.cpp prebuilt-CUDA install location, the vLLM/SGLang
  Linux-only note, and the LOCAL_STUDIO_CHROME_PATH browser override.

Docs-and-preflight only; no runtime code changes. The Windows llama.cpp install
and browser behavior it documents ship in the wave-3 PRs (llamacpp-install,
browser-host).

Validation

bash -n scripts/doctor.sh clean. Ran npm run doctor live on Windows 11
(Git Bash): all three new Windows branches fire correctly —
PASS python3 3.11.9 found, INFO nvidia-smi found — CUDA llama.cpp serving available (vLLM/SGLang require Linux), PASS models dir C:\Users\…\models exists and is writable — 7 pass / 0 fail. On Linux the script takes the unchanged
POSIX branch.

UI changes

None.

Risks / rollout notes

• Pure docs + shell-preflight change; no effect on the controller, frontend, or
  CI beyond the doctor script's own output.

Fixes #214.