docs(wsl): add end-to-end LIBERO eval on WSL2

[VatsalyaBhadaurya]

What this adds

Follow-up to #12. Documents running the full serving stack (vla-server + LIBERO client) inside a single WSL2 distro, so the complete loop — sim reset/render → client preprocess → vla-server inference → action → sim step → video — runs on one Windows machine with no separate Linux host.

• New "End-to-end evaluation (LIBERO client on WSL)" section in docs/backend/wsl.md, plus a SmolVLA / RTX 4050 result table.
• New eval/run_libero_wsl.sh — a one-command runner that starts vla-server, waits for ready, drives one client run, and stops the server on exit. Takes task-id and episode count as args, or prompts for them; model/suite/output overridable via env.
• Drops the stale patches/patch.sh reference (CMake fetches llama.cpp automatically now).

Key WSL-specific findings

• Work from the Linux filesystem, not /mnt — uv/venv installs are unreliable on drvfs; clone into ~ and run there.
• Add the user to the render group so MuJoCo EGL can open /dev/dri/renderD128 instead of falling back to software rendering (Permission denied -> kms_swrast).
• Set UV_LINK_MODE=copy + a home-dir cache so uv does not hardlink across filesystems.

Usage

bash eval/run_libero_wsl.sh          # prompts for task-id + episodes
bash eval/run_libero_wsl.sh 3 5      # task-id 3, 5 episodes, no prompt

Verification

Ubuntu 24.04 (WSL2) + CUDA 12.6 + RTX 4050, server + cl

- vla: backend = CUDA (device 0), ~1.07 GB VRAM
- libero_object task 0 succeeded end-to-end: is_success
- ~131 ms/req server-side (vision ~84 / inference ~46), ~155 ms/step client-side incl. software EGL rendering
- Rendered episode_000000.mp4 + summary.txt produced
- eval/run_libero_wsl.sh 0 1 reproduces the run: CUDA server up → episode succeeds → server cleaned up

Smoke episode

wsl_smoke_episode.mp4