Move metrics and probe listeners from main.go into server.go and
manage all servers with errgroup. This replaces log.Fatalf goroutines
with proper error propagation and ties all server lifetimes together
so a failure in any listener triggers graceful shutdown of the others.
Without IdleTimeout, clients holding keep-alive connections open indefinitely
caused server-side goroutines (and their ~16KB of buffers) to accumulate
linearly until OOM.
Reduces IdleConnTimeout to 5 seconds. Reverse proxying to a server that
has a shorter keep-alive may cause "EOF" and "connection reset by peer"
issues as the connections may be closed by the upstream before our
client notices.
