sdk/python: keep a handle's relay connection across calls

[CMGS]

Fourth step of #195: the Python SDK side of a persistent data-plane connection, the same shape as #198 for Go. Needs #196 in the guest image to take effect; against an older silkd it behaves exactly as before.

What changes

• A Sandbox handle keeps its relay connection. After a call the connection is parked in a ConnPool (up to 8 per handle, 30 s idle; Client(..., keep_alive=...) tunes the window, 0 dials per call) and the next call sends its request on it. Every RPC helper runs under _park_after: a terminal frame, error frames included, parks the connection; a dropped connection, a timeout cut or a local error drops it. close and hibernate drain the pool first.
• What is not reused. watch, open_pty, dial_port and an LSP session own their connection as today. run joins its stdin thread before the connection goes back, so the pump's frames never land on the next RPC (silkd drains what an exited command left unread).
• Liveness before reuse. A parked connection is checked with a zero-timeout select: a peer that hung up (hibernate, release, the idle sweep) or spoke unprompted fails it and the call redials, which wakes the guest exactly as before. It works through TLS sockets too, since it never consumes bytes.
• Protocol negotiation. A handle's first dial asks the daemon's proto with one info round trip. A daemon before proto 2 closes after answering, so the handle redials and dials per call from then on. One extra round trip once per handle.
• logs writes exit then done; the trailing done is now consumed so the connection stays in frame. docs/sdk-python.md documents the window and its interplay with idle_hibernate_seconds.

Hot path

A pooled call: one lock pair and one select(2), replacing a TCP dial, an HTTP upgrade, a TLS handshake behind an edge and a vsock connect on the node. The first call on a handle pays one info round trip. Frames on the wire are unchanged.

Evidence

$ ruff format --check . && ruff check .
33 files already formatted
All checks passed!
$ (cd sdk/python && python -m pytest -q)
193 passed in 28.02s
$ (cd sdk/python && python -m mypy)
Success: no issues found in 9 source files
$ (cd sdk/python && for i in 1 2 3 4 5; do python -m pytest -q tests/test_keepalive.py | tail -1; done)
6 passed ×5
$ (cd sdk/langchain && PYTHONPATH=../python python -m pytest -q)
6 passed in 0.16s

Run on Python 3.9.6, the declared floor. The OpenAI adapter suite could not be collected locally (no agents package in this environment); CI runs it on 3.12.

New tests in tests/test_keepalive.py against an in-process fake agent that answers info, fs_stat and exec: four calls share one upgrade; an old daemon costs the probe dial plus one per call; keep_alive=0 dials per call; an idle connection closes after the window; a peer that hung up is noticed before reuse; close drains the parked connection. The fake half-closes and drains like the relay does, since a plain close with an unread stdin_close resets the reply. Scripted one-shot fakes in the existing tests run with the proto pinned to 1, the fallback path; the logs script now ends exit, done as silkd does.