WIP: Add a scenario guide for deploying advanced serving techniques

[dennis-upbound]

Description of your changes

Modelplane's pitch is that adopting an advanced inference-serving technique should
be a deployment-level change, not a quarter-long project — but we have no
end-to-end, customer-facing walkthrough that shows it. A prospective ML team can't
see how they'd evaluate something like prefill/decode disaggregation on their own
workload and roll it out to production safely.

This adds a self-contained scenario guide and runnable kit under
demo/prefill-decode/. The guide (README.md — "Deploying advanced serving
techniques") opens with the latency problem (TTFT/ITL under load), explains why
these techniques are powerful but operationally heavy to ship, then walks the real
adoption workflow with prefill/decode disaggregation as the worked example, on one
Modelplane cluster:

• a unified baseline ModelDeployment → a PrefillDecode variant stood up
  as a second deployment on spare capacity → proving it genuinely disaggregates
  via vLLM's own NIXL KV-transfer counters → benchmarking a replay of
  representative traffic → promoting it behind a single shared-label
  ModelService by shifting replica capacity. Because the service load-balances
  evenly across healthy endpoints (one endpoint per replica), canary → promote →
  cut over → roll back are all just replicas — no traffic weights, no platform
  ticket.

It's backed by runnable artifacts, not just prose: manifests/ (cluster, cache,
unified + P/D deployments, the shared ModelService), a single
run.sh {deploy|prove|bench|promote|rollback}, and a replay-trace.jsonl. The
disaggregation recipe matches what #175 validated live — vanilla
vllm/vllm-openai:v0.19.1 (ships NIXL), kv_producer/kv_consumer, decode
--port=8001, matched --block-size — with Modelplane composing the EPP /
InferencePool / routing sidecar / NIXL plumbing.

Draft / WIP: the guide's benchmark curve and offload counters are placeholders
pending a live run on a fresh cluster. I'll fill the real numbers and confirm the
GuideLLM --data replay shape against the live endpoint, then mark ready.

Reviewers — the judgment calls are the narrative framing (does the problem →
technique → scenario arc land for an ML-team audience?) and whether the
promotion-by-replica-count story reads as faithful for a production rollout.

Builds on #175.

I have:

• Read and followed Modelplane's contribution process.
• Run nix flake check (or ./nix.sh flake check) and made sure it passes.
• Added or updated tests covering any composition function changes. (docs, manifests, and scripts only — no composition function changes)
• Signed off every commit with git commit -s.