runtime: allocate unique event IDs across wraparound

[AFOliveira]

Problem

EventId is a 16-bit value. EventManager::getNextId() incremented through wraparound and ignored whether std::set::emplace() actually inserted the ID. If an older command with that ID was still in flight, the newer command reused it; the first completion erased the shared entry and the second produced Couldn't dispatch event: 65535. Long ET inference runs reproduced this and could leave the device unhealthy.

Fix

• scan for a free ID at wraparound and never reuse an in-flight or still-blocked ID
• fail before submitting a duplicate command if all 65,536 IDs are concurrently occupied
• retain the ID until runtime observers consume the completion
• add wraparound, collision, and exhaustion regression tests

Validation

• git diff --check
• focused tests are included; full container build/CI is pending on this draft

Paired host/CI containment is being prepared in aifoundry-org/hf-hackathon#117.