fix: harden selective retransmit test with multi-attempt strategy

[HMBSbige]

The dtls13_selective_retransmit_only_missing_records test was flaky because it assumed the first retransmission after a dropped packet would always be selective. In practice, a duplicate-triggered (dupe triggers resend) full-flight resend can occur before the peer's ACK narrows the retransmission scope.

This PR rewrites the test to be robust against that timing:

• Run up to 12 attempts with varying deterministic RNG seeds and drop positions, so different fragmentation layouts are exercised without relying on runtime RNG.
• Deliver all client packets normally (no epoch-0 filtering), keeping the handshake state machine realistic.
• Tolerate full-flight resends, and keep waiting until a selective resend (fewer packets than original flight) is observed.
• Require both selective retransmit and successful handshake completion before declaring success.
• Report detailed diagnostics on failure (attempts with drop / retransmit / client epoch-2 delivery / connection counts).

[algesten]

Thanks!