fix: keep TCP tunnels recovering under sustained traffic

[dr-hoseyn]

Under sustained traffic, stale generation workers, undercounted concurrent pool dials, retained mux claims and unbounded control/destination writes could stop recovery until restart. Raw UDP also expired active one-way flows, truncated large datagrams and could panic when the opening packet raced queue shutdown.

Close and join generation work before replacement, replenish and accurately account for pools, release mux capacity on every exit, drain retired queues and bound handshake/control/destination writes. SMUX growth respects the initial pool and a 128 MiB aggregate receive-window budget. Framing handles short writes and rejects oversized frames. Pooled WebSocket pings use the existing ten-second write bound, and TCP-to-WebSocket copies preserve bytes returned with EOF.

Raw UDP now tracks activity in either direction and joins both copy workers on failure. Queue the initial payload before publishing a flow. Full-sized receive buffers preserve legal large and empty datagrams; each packet is written once, and a short write ends the flow rather than splitting it. Larger raw buffers increase memory per flow; framed UDP's adaptive buffers remain separate.

Guard disabled per-packet Debug logging. Same-runner measurements reduce client UDP logging overhead from 112 B / 10 allocations to 4 B / 1 allocation per packet, and server copy directions from 64 B / 5 allocations to zero. Exact-source races and paired measurements pass at b7c23a6. Separate actual-tunnel checks cover delayed replies after 60.5 seconds of continuous one-way traffic, 16,384/16,385/60,000/zero-byte packets and repeated shutdown admission.

Carries forward the lifecycle/pool intent of #17, #19 and #20 on current main. One earlier KCP churn goroutine-count failure did not reproduce in the final full suite or five isolated race repetitions.

Validation: Combined full tests, full race suite and unchanged coverage gates pass, including five separate KCP churn race repetitions. Final integration with UDP backend health passes affected runtime races, vet, staticcheck and Linux amd64/arm64 builds. The full-suite run precedes the small UDP backend health update; the final integration validates that update.

Only existing production and test files change. No files, test files, dependencies or configuration/wire-format changes are added. Supplemental fixtures and CI workflows remain outside this PR on fork validation branches. Measurements use local test traffic and do not establish WAN throughput.