Add Pi ACP devcontainer with local-LLM + jupyter-mcp support and E2E test

[Copilot]

Adds a Pi coding-agent devcontainer wired through pi-acp, preconfigured for local LLM providers and MCP servers (with Datalayer's jupyter-mcp-server baked into the default config), plus an end-to-end test that exercises the kernel against the real pi-acp binary.

Devcontainer (.devcontainer/pi/)

• Dockerfile: Node 22 + @earendil-works/pi-coding-agent + pi-acp + pi-mcp-adapter installed under ~/.local as jovyan. Follows the Codex pattern of COPY . /home/jovyan/, so the project's .pi/agent/*.json lands at ~/.pi/agent/*.json and is picked up as pi's user-global config.
• devcontainer.json: ACP_AGENT_COMMAND=/home/jovyan/.local/bin/pi-acp, secrets for Anthropic/OpenAI/Google/DeepSeek, and pass-through env for LM_STUDIO_BASE_URL / OLLAMA_HOST. PI_ACP_ENABLE_EMBEDDED_CONTEXT=true.

MCP defaults (.pi/agent/mcp.json)

Routes MCP through pi-mcp-adapter (the supported path — pi-acp itself does not forward ACP-supplied MCP servers to pi, per its README):

{
  "mcpServers": {
    "jupyter": {
      "command": "uvx",
      "args": ["jupyter-mcp-server@latest"],
      "lifecycle": "lazy"
    }
  }
}

End-to-end test (tests/e2e/test_pi_acp_integration.py)

• Spawns real pi-acp, drives it through the kernel's ACPClientImpl + connect_to_agent, and asserts initialize + session/new complete (the latter forces pi-acp to spawn pi --mode rpc successfully).
• Auto-skips via shutil.which when pi/pi-acp aren't on PATH, so the default unit-test matrix is unaffected.
• New pi-acp-e2e CI job (Node 22 + global install of pi/pi-acp) runs it.

README

Adds the Pi devcontainer to the list with notes on local-provider and jupyter-mcp-server support.