The hardcoded 128 KiB cap caps single-flow upload throughput at roughly
value / RTT on high-BDP links (~10 Mbit/s at 100 ms RTT), manifesting
as slow uploads through the proxy. Expose it as
[network] tcp-not-sent-lowat so operators can raise it on fast links,
while keeping the default behaviour unchanged.

TCP keepalive was configured (SetKeepAlivePeriod) but never actually
enabled (SO_KEEPALIVE) on accepted client connections. Go 1.26's
SetKeepAlivePeriod only sets TCP_KEEPIDLE — it does not call
setsockopt(SO_KEEPALIVE, 1). Without SO_KEEPALIVE the kernel never
sends probe packets, so dead connections from sleeping mobile clients
linger until the idle timeout fires.

Replace SetKeepAlive + SetKeepAlivePeriod with net.KeepAliveConfig
(available since Go 1.24) for explicit per-socket control:

  Idle:     30s   (time before first probe)
  Interval: 10s   (between probes)
  Count:    3     (failed probes to declare dead)

This detects dead connections in ~60s instead of relying on system
defaults (tcp_keepalive_intvl=75s, probes=9 → up to 11 minutes).

Increase the default idle timeout from 1 minute to 5 minutes.
MTProto clients send ping_delay_disconnect every ~60s, which resets
the idle timer. The previous 1-minute default created a race: if a
ping arrived even 1–2 seconds late the relay was killed. A 5-minute
window also survives typical mobile sleep periods (phone idle 2–5 min)
where the NAT mapping is still alive and the connection can resume
without reconnection.

Ref: #132