fix: reap completed response send tasks

[N0zoM1z0]

Summary

• reap completed response-send tasks during the main service loop
• preserve concurrent response writes and the existing shutdown drain
• log a join failure when a response-send task panics

Why

response_send_tasks currently retains every completed task until transport shutdown. A long-lived stdio server therefore grows resident memory with each request even though the response write has completed.

In a sequential 5,000-call stdio soak, one downstream server grew from 42.9 MiB RSS after warm-up to 60.3 MiB. Direct calls to its handler stayed flat, isolating the growth to the SDK serve loop. With this change, the same stdio workload stayed between 46.1 and 47.3 MiB.

Verification

• cargo +1.97.0 fmt --all -- --check
• cargo +1.97.0 check -p rmcp --features server,transport-io
• cargo +1.97.0 test -p rmcp --features server,transport-io --test test_stdio_response_concurrency

[DaleSeo]

Thanks for the fix, @N0zoM1z0!