Add Docker Compose evaluation stack (client image + orchestration + docs)

[AIWintermuteAI]

Description:
This PR adds a Docker Compose evaluation workflow for running LIBERO /
SimplerEnv simulation episodes against the vla.cpp inference server.

What's included

eval/Dockerfile.client — Python simulation client container

• Based on ubuntu:22.04 with Python 3.10
• Installs MuJoCo 2.3.x, robosuite, lerobot, and all eval dependencies
• CPU-only PyTorch (server handles GPU inference)
• LIBERO installed via editable mode (keeps data files accessible)
• All dependency versions pinned to avoid known incompatibilities
  (NumPy 1.x, MuJoCo 2.x, Pandas 2.0.x)

eval/docker-compose.yml — Two-container orchestration

• server service: builds from the upstream Dockerfile (CUDA)
• client service: Python simulation environment
• Shared volumes for model files and eval outputs
• CDI-based GPU access (no nvidia-container-toolkit required)
• Default command launches SmolVLA for LIBERO

docs/DOCKER.md — Workflow documentation

• Quick start guide (download models → build → run evaluation)
• Configuration reference (volumes, ports, network, CDI)
• Running without Docker Compose (standalone docker build/run)
• Known issues table with workarounds

Usage

docker compose -f docker-compose.yml build
docker compose -f docker-compose.yml up -d server
docker compose -f docker-compose.yml run --rm client \
    python run_sim_client_direct.py \
        --task libero_object --task-id 0 --n-episodes 1 \
        --output-dir /tmp/libero_outputs --arch smolvla \
        --vla-addr tcp://server:5555

Validation

Tested end-to-end on RTX 5060 Ti (CUDA arch 120) with SmolVLA LIBERO:

• 220-step episode completed successfully
• Reward = 1.0, 116.45 ms average inference per step