Cut fixed per-request costs out of the reshard control plane

[yinlin09]

Every Stage-3 coordination pays fixed costs on its framed RPCs (coordinate, GET_METADATA, receiver arm):

1. The 4-byte length prefix and the body go out as two separate send() calls with Nagle enabled, on requests and responses alike, exposing every hop to the delayed-ACK stall.
2. The destination controller is asked for every registered unit's full pool manifest on every request, although work units register once per engine lifetime.
3. Client sockets carry no keepalive, so a black-holed peer is only detected at the full receive timeout (300 s on the coordinate hop).
4. The framed server never reaps its per-connection threads: one std::thread handle and stack per request, held until shutdown.

Changes: single-buffer framing both directions plus TCP_NODELAY on client and accepted sockets; SO_KEEPALIVE and TCP_USER_TIMEOUT bounded by the call's I/O timeout; the accept loop joins finished connection threads; and the destination metadata is cached per controller address. Staleness (engine replacement) surfaces as a plan-build or receiver-arm failure. A failed attempt that used the cache drops the entry; while no receiver has acknowledged its arm the attempt is side-effect-free beyond the abandoned claim and is replayed once on fresh metadata. Once any receiver has acknowledged (multi-destination plans arm every receiver concurrently), the failure is returned as is and the next request re-queries.

Validation

reshard_service_test passes in the OSS bazel build (ml-build container, clang-18), including the two new tests: RemoteMetadataCachedAndRefreshedOnStaleFailure (cache hit on the second request; exactly one refetch after a fingerprint-mismatch replay) and PartialReceiverArmFailureDropsCacheWithoutReplay (one of two receivers refuses its arm: no replay, the cache entry is dropped, the next request re-queries).

Benchmarked on tpu7x 1P1D (Qwen3.5-397B, prefill PCP8 → decode DP8, vllm-torchtpu with background Stage-3 submission, 888 transfers per leg, zero transfer failures either leg): plan build including the peer-metadata path drops from p50 2.1 ms / max 5.1 ms to p50 1.5 ms / max 3.0 ms; end-to-end serving metrics are unchanged. The coordinate call's floor is dominated by sender dispatch (p50 ~26 ms under 4-way concurrent coordination — the reply currently awaits all 8 sender-worker acks), which this PR deliberately does not touch: returning the coordinate reply at receiver-arm and letting sender dispatch complete asynchronously is the follow-up that would remove most of the remaining floor.