Mooncake#4209: TCP transport queue-full after idle gap causes PD hang

TCP transport: first KV transfer after idle gap (~5s) fails with queue-full rejections. Producer rejects all slices, batch_transfer_sync_write returns -1. Consumer hangs until timeout. Self-heals after ~90s. Root cause: retired group not removed from state->groups. Demonstrates multi-GPU silent failure in KV transfer.