Add MI355X hybrid GDN profiling and SGLang correlation

[powderluv]

Summary

• Add ROCm MI355X device and Qwen3.8 hybrid GDN model contracts.
• Add ROCm-aware profiling support for GDN, attention, dense, routed-MoE, and collectives.
• Add strict SGLang capture, routing, graph-event, replay, and exact-cost validation tooling.
• Document reproducible workflows while keeping generated traces, profiles, captures, and benchmark reports external.

Validation

• 411 changed-file tests passed; 2 GPU-dependent tests skipped in the local suite.
• Additional ROCm-host coverage passed for tests requiring PyTorch/ROCm.
• Diff formatting, compilation, runtime identity, exact query coverage, rank coverage, and perturbation gates were checked.

Scope

The runtime-cost validation is limited to one-node TP-only non-speculative decode. It does not claim serving, expert-parallel, multi-node, or full-forward parity. No generated performance results are included in this PR.