Handle unchecked I/O failures in connectivity validation

[kasnder]

The connectivity probe catches IOException, but Android's socket factory can throw UncheckedIOException when duplicating a descriptor fails with EMFILE. That bypasses the normal failure backoff even though the current executor captures the exception.

Extract the socket operation into a small helper that converts unchecked I/O failures to IOException, preserving the complete exception chain. The existing probe failure handler then applies its normal backoff. Keep the network-bound socket factory, port 443 and 10-second connect timeout; use try-with-resources for returned sockets. Unrelated runtime exceptions still propagate.

Validation: 11 focused tests passed (NetworkValidationProbeTest and ServiceSinkholeValidationBackoffTest), covering checked and unchecked allocation/connect failures, socket cleanup, endpoint/timeout and existing backoff. Tests use a numeric address and fake sockets without network access. :app:lintGithubDebug passed with no errors; git diff --check passed.

Related to #932. This fixes probe failure handling, not the unidentified cause of long-uptime descriptor exhaustion. Current Android 16 source already closes a factory socket if binding fails, so this PR makes no factory-leak claim. No device exhaustion reproduction; keep the issue open for that investigation.