fix: allow fragmentation for inbound UDP replies

[pengwk]

Summary

• honor the existing outbound_udp_allow_fragmentation option on inbound UDP listener sockets
• propagate global and per-server settings, including manager mode
• preserve the current default and cover Unix and Windows
• add a Linux regression test for IP_MTU_DISCOVER

Problem

The UDP relay sends encrypted responses back to clients through its inbound listener socket. That socket always enabled DF, even when outbound_udp_allow_fragmentation was set. A near-MTU target response can grow after encryption and then fail with EMSGSIZE on the client-facing send path.

This change reuses the existing option instead of adding another configuration knob, making it apply to both UDP relay directions.

Verification

• Linux: cargo test -p shadowsocks --test udp --no-default-features --features aead-cipher-2022 (4 passed)
• Linux: cargo check -p shadowsocks-service --no-default-features --features server,aead-cipher-2022
• Real relay path: a 1472-byte UDP response failed before the change and passed 10/10 after it with fragmentation enabled