Add stream-aware device overlap verification tests

[nicksunka]

Summary

Adds focused regression coverage for stream-aware accelerator overlap verification.

The existing verify_kernel_overlap() implementation already handles accelerator events generically through the Torch accelerator event category and separates verification queues using pid and args.stream.

These tests explicitly verify the expected stream-aware overlap behavior for kernel and memory events without introducing any new overlap detection logic.

Coverage

The tests verify:

• Same-stream kernel/kernel overlap is detected.
• Cross-stream kernel/kernel overlap is allowed.
• Same-stream kernel/memory overlap is detected.
• Cross-stream kernel/memory overlap is allowed.
• Same-stream memory/memory overlap is detected.
• Cross-stream memory/memory overlap is allowed.
• Nested same-stream memory overlap is detected.
• Back-to-back same-stream memory events are allowed.
• Events without a stream are assigned to the same default stream for verification.
• A temporary default stream added during verification is removed from the original event afterward.

Validation

Focused overlap verification tests: 10 passed

Full AIU Trace Analyzer test suite: 215 passed, 5 xfailed

No production behavior is changed by this PR.

Related to #122

[lasch]

lgtm