fix(net): preserve TCP port ownership through TIME_WAIT

[fslongjin]

Summary

Fix TCP port ownership and TIME_WAIT reuse rather than releasing all protection when a descriptor closes.

• Track FD bindings and protocol lifetimes separately under one namespace-wide port authority, with indexed tuple ownership and independently inherited accepted-child state.
• Apply both-sided SO_REUSEADDR rules, atomic listener promotion, and explicit nonzero-port retention on connection failure. Release unlocked bindings on protocol termination even when the old descriptor remains open.
• Transfer unreachable TIME_WAIT sockets into compact protocol storage instead of discarding them or retaining full send/receive buffers. Keep expiry and namespace routing active in direct and NAPI polling.
• Enable per-connection TCP timestamps and 60-second TIME_WAIT protection. Refresh the monotonic context after acquiring protocol locks so idle active-reuse attempts can advance without unrelated traffic.
• Pin the smoltcp implementation and add 31 dunitest cases to the mandatory suite, including isolated two-port automatic reuse and unconsumed nonblocking-refusal close paths.

Dependency: DragonOS-Community/smoltcp#33 is merged into dragonos/v0.12.0. Pin its merge commit 9b9813e25323fb574f80b8aebfbbf2e7809f47ab; its Git tree is identical to the previously validated revision.

Validation

• make fmt, make kernel, and whitespace checks passed.
• DragonOS/QEMU: 31 new cases and 168 existing network cases passed.
• Linux comparison: 31 cases passed; the original bind/child/TIME_WAIT reproducer now agrees with Linux.
• smoltcp: 671 library tests and seven feature configurations passed; isolated port-logic tests: 8 passed.
• nginx: 10 rounds of dual-stack HTTP, graceful exit, and immediate restart passed.
• 40,000 client-active closes and 40,000 server-active closes passed, approximately 18.8s and 19.1s respectively.
• A real 60-second protection interval expired with the original FD either open or closed (60.177s/60.176s); the two-port reconnect test requires no auxiliary packets.
• 1,024 additional unconsumed nonblocking-refusal closes did not exhibit a full-buffer-per-close memory leak.
• Independent security/concurrency, correctness, and resource/performance reviews were completed; confirmed findings were fixed and re-reviewed.

The 2,000-connection median changed from 905.7ms to 925.9ms (about +2.2%) while retaining the previously discarded protocol protection. Fixed-server-port batches remained approximately flat through 40,000 historical closes. Peak memory includes retained protocol records and allocator/container capacity; this is not a zero-cost compatibility change.

Scope

Dynamic migration of the same local IP between interface-owned protocol stacks remains a separate issue. Ordinary route/egress changes are not that migration case. This change does not relax tuple protection to hide that limitation and does not add a general TCP sysctl or recoverable-allocation framework.

[fslongjin]

@codex review