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.

Fixes #439.

When `[network] dns = "tls://..."` (or "https://...") is set, the
resulting *net.Resolver gets attached to the base network's NativeDialer
and was previously also handed to golang.org/x/net/proxy.FromURL via
NewProxyNetwork. As a result, the SOCKS5 client used the user's DoT/DoH
resolver to look up the SOCKS server's own hostname (e.g. "xray" inside
a docker compose stack). Public DNS-over-TLS resolvers don't know about
docker-compose service names, k8s service DNS, /etc/hosts entries, or
corporate split-horizon DNS, so the upstream lookup returned NXDOMAIN
and the proxy chain broke with a misleading "lookup xray on
127.0.0.11:53: no such host" error.

The custom DNS resolver exists to bypass DPI poisoning when resolving
public censored names like Telegram DCs or the SNI/fronting host. Proxy
server addresses are almost always internal and should be resolved via
the system resolver instead. This change introduces proxyServerDialer,
which copies the timeout and fallback-delay from the base dialer but
leaves Resolver==nil, and uses it for the SOCKS upstream.

The new internal test asserts the structural property directly: the
returned dialer must not inherit the base's custom resolver.