sdk/python: dial routers and DHT peers at once, stop when the record is in hand

[aojea]

The bananas deploy of bd58efb failed the SDK cold-path probe again (https://github.com/google/sam/actions/runs/36338262423/job/108677106672); with #532's timeout the probe now reports it instead of hanging:

{"probe":"sam-sdk-python","ok":false,"stage":"node-mcp","error":"no answer with 'Echo: sam-sdk-probe' in 120s (1 attempts; last: on the mesh as 12D3KooWPZ… (hung for 120s))"}

Measured in the probe's pod (same image, token and env)

Cold path before: enroll 0.2s, join 16s, discover 15s, connect 0.2s. Each 15s is one bounded dial to the GCE router the control plane lists but the cluster's network policy drops — and every step dialed one peer after another. Right after a rollout the routers' tables also name the pods just replaced as closer peers (another 15s each), which is what pushed the first attempt past 120s on the deploy and made the hub's first attempt a coin toss.

After, three cold runs: join 15.3s (the dark router's timeout, alone), discover 0.1s, connect 0.3s, call 0.2s.

Changes

• One shared DHT walk (_walk) for find_peer/find_providers: a round asks its peers at once and returns as soon as an answer satisfies the lookup; only otherwise are the closer peers dialed, also at once. A round costs one query and one dial timeout, not one per peer. Drops the duplicate GET_PROVIDERS helper.
• _admit dials and authenticates every router at once; the reservation still goes to the first admitted router of the list (same reachability semantics as before).
• Tests: providers in hand end the walk with no dial; three stale closer peers cost one dial timeout together; a slow router does not delay the others; join with a dark router first in the list ends at DIAL_TIMEOUT with the reservation on the first admitting router.

Validation

• sdk/python: 91 passed
• go test ./tests/integration -run 'TestNativeSDK|TestSDK': all pass (incl. TestNativeSDKsAcrossRouters, JS and Python)

[gemini-code-assist]

Code Review

This pull request refactors the Kademlia walk in discovery and the router admission process in session management to perform network queries and dials concurrently using Trio nurseries, significantly reducing the impact of slow or unreachable peers. New tests have been added to verify these concurrent behaviors. The feedback suggests refactoring a list comprehension inside an any() call to prevent potential short-circuiting bugs if it is ever refactored into a generator expression, ensuring all provider answers are always accumulated.