sdk/python: a peer that is gone costs connect one wait, not one per router

[aojea]

The sam-sdk-python cold-path probe failed on the bananas deploy of c916264 (and in four scheduled runs before it, on the previous image), always at the node-mcp stage as hung for 120s with 1 attempts, while the JS probe passed. This is what the Python SDK did with the time.

After a rollout the routers' tables still name the pods it replaced. connect() walked the relayed paths in turn: a relay whose destination is gone answers a CONNECT only after its own 30s timeout (ConnectTimeout in go-libp2p's relay), and a dead provider was tried through every router twice, once as an admitted router and once more as the relay the DHT named for it. py-libp2p's own limits on a relayed connection's upgrade add another half minute or more when the relay accepted the circuit and the far end never speaks. The first stale provider alone used up the probe's budget.

• session.py: the routers of each step are dialed at once, as join() and the DHT walk already do, and the first circuit that opens ends the others. The DHT step skips relays that are admitted routers, which the first step dialed. The JS SDK dials every address of a peer at once already.
• relay.py: a relayed dial is held to DIAL_TIMEOUT as a direct one is, and a circuit given up on (timed out, failed, or cancelled because another path won) is reset so the relay drops it too.
• mesh.py (first commit): the refresh loop and the control plane pull run in separate worker threads and both persist the credential with the same temporary file name; one probe run logged [Errno 2] No such file or directory: 'identity.key.tmp' -> 'identity.key'. The control plane also redeems only the last biscuit it issued, so two refreshes at once leave the loser with a spent one. A lock on the member serializes refresh, the pull and save.

Tests:

• test_session.py: three routers that all wait on a dead destination cost connect() one wait; one that opens ends the others' attempts.
• test_relay.py (new): a relay that answers CONNECT with OK and forwards nothing; the dial ends at DIAL_TIMEOUT and the relay sees the circuit let go. Hangs past 20s on the previous code.
• test_mesh.py: the fake control plane redeems only its last biscuit, as the real one does; eight concurrent refreshes all land. Seven fail with biscuit already redeemed on the previous code.
• TestNativeSDKsMesh gains a stalled-relay step: a Go relay that accepts every circuit and forwards nothing, dialed by every SDK member at once. Each gives up within its dial timeout and the relay counts every circuit let go. The Python member exceeds the 20s budget on the previous code.

make sdk-python (96 passed), TestNativeSDKsMesh and TestNativeSDKsAcrossRouters pass with both SDK toolchains present.

[gemini-code-assist]

Code Review

This pull request introduces concurrency improvements and timeout boundaries to the Python SDK and integration tests. Specifically, it adds thread-safety locks to credential refresh and control plane synchronization in mesh.py, implements concurrent dialing of routers in session.py to prevent sequential timeouts, and bounds relayed dials with a timeout in relay.py. It also adds corresponding unit and integration tests to verify these changes under concurrent and stalled conditions. There are no review comments to address, and the changes look solid.