feat: multi-GPU topology with NVLink bandwidth model

[manishklach]

Summary

Expand the topology model to represent more realistic multi-GPU fabrics with explicit NVLink bandwidth limits, asymmetry, and contention behavior.

Scope

• add richer GPU-to-GPU path modeling
• differentiate local NVLink islands from PCIe fallback paths
• surface the effect in placement, movement, and congestion metrics
• add tests and one benchmark scenario focused on intra-node topology

Why

The repo already treats KV as a distributed systems problem. A more faithful intra-node fabric model would make locality decisions and movement costs more believable.