fix: correct UDP errqueue classification and timerfd polling tests

[fslongjin]

Summary

Address two independent CI failures:

1. UDP error queue: classify EMSGSIZE reporting by the actual packet destination, not only the socket's creation family. IPv4 and IPv4-mapped sends must not produce native IPv6 local error records. Reuse the resolved destination snapshot for connected sends on the ordinary output path.
2. timerfd regression test: equal deadlines do not guarantee that one poll observes all expiry callbacks. Accumulate unique ready descriptors under a fixed monotonic deadline while preserving cancellation, tick-count and cleanup checks.

The UDP policy remains in the existing error queue helper. There are no new locks or protocol states, no changes to timerfd kernel scheduling, and no modifications to gVisor reference tests. The correct IPv6 socket name-query ABI is preserved.

Root cause and coverage

• Original CI failures: dual-stack PayloadTooBig/2 and /5 returned an error queue record instead of EAGAIN; the same-deadline timerfd test observed 9 ready descriptors while requiring 16 in one call.
• Added UDP coverage spans IPv4, IPv6 sockets with IPv4 sockaddr, mapped and native IPv6 destinations, connected/explicit sends and RECVERR enablement. Positive cases validate error metadata, destination address/port, and queue consumption.
• The timerfd case still requires all 16 active timers, exactly one tick each, no cancelled timer events, and empty descriptors after consumption. It does not weaken the expected count or rely on extra sleeps.

Validation

Environment	Result
Unmodified master in DragonOS	Original gVisor dual-stack failures and both new UDP regression failures reproduced
Linux	UDP dunitest 22/22; timerfd 9/9; same-deadline case 20 iterations; gVisor PayloadTooBig 6/6
Fixed DragonOS, QEMU/KVM x86_64, 2 vCPUs	UDP 22/22; timerfd 9/9; poll 4/4; alarm 4/4
Fixed DragonOS repeated timerfd case	30 additional iterations passed
Unmodified full gVisor UDP loopback binary	528 passed, 18 existing skips, zero failures
Build / formatting	make kernel; FMT_CHECK=1 make fmt; git diff --check

Independent plan, UDP implementation and test reviews found no blocking issues.

Limits

The old timerfd assertion passed ten local baseline iterations; its failure was captured in CI and its assumption is not guaranteed by poll semantics. This change does not claim complete MTU/IPv4 ICMP error queue compatibility. Native IPv6 non-loopback output was reviewed at the code level, not separately exercised end-to-end. Full repository-wide gVisor and dunitest runs are left to CI; existing skips are not counted as passes.

[fslongjin]

@codex review