Build inference request router for distributed forward pass

[EightRice]

Goal

When a user inference request lands on a node, route it across the peers that host the model's modules, orchestrate the forward pass, and return the result.

Scope

• Define the request envelope (input data hash, requested model hash/manifest, return-to address).
• Look up which peers host which modules (via P2P capability gossip — already exists in p2p.py).
• Plan a forward-pass execution path across peers (linear pipeline first; tensor parallelism later).
• Drive the tensor-transport protocol step by step.
• Handle peer dropout gracefully (pick alternate host or fail with a clear error).
• Emit the events that RPB.recordInference() listens for (provider, requester, units, normalized cost).

Depends on

• Tensor transport over libp2p (separate issue)
• A loadable trained-model manifest (separate issue, currently a single-CID stub on RPB.setMatureModel)

Acceptance

• One end-to-end test: user submits an inference request → router orchestrates → result returns → InferenceServed event fires → recordInference accounting matches expectation.