Feature: GPU placement simulator for pipeline rollout planning

## Summary
Add a GPU placement simulator that predicts MIG slice usage, latency risk, and admission outcomes before operators activate new streams, models, or pipeline revisions.

## Problem / Opportunity
The project has a GPU scheduler, MIG defaults, canary/shadow workflows, and cost accounting, but operators lack a dry-run planning tool. Today, a rollout can be syntactically valid yet operationally risky because the expected GPU slice mix, concurrency, fallback behavior, or tenant contention is not visible until execution.

## Proposed Feature
Expose a simulation API and console workflow that takes proposed streams, model versions, pipeline revisions, and tenant priorities, then returns expected placement, resource pressure, cost estimate, and admission/refusal reasons without changing live scheduler state.

## Scope
- Add scheduler dry-run inputs for candidate workloads and time windows.
- Model MIG slice availability, current reservations, tenant priority, and known model resource profiles.
- Return placement candidates, bottlenecks, and refusal explanations.
- Integrate with pipeline/model promotion and stream creation flows as an optional preflight.
- Add metrics and docs for interpreting simulation confidence.
- Add tests with synthetic cluster topologies.

## Acceptance Criteria
- [ ] Users can dry-run a candidate pipeline/model/stream rollout against current scheduler state.
- [ ] Results include projected MIG allocation, unresolved constraints, and estimated incremental GPU cost.
- [ ] The simulator distinguishes hard admission failures from soft risk warnings.
- [ ] No live reservations are created during simulation.
- [ ] Console rollout screens can invoke simulation before activation.
- [ ] Documentation describes simulator assumptions and failure modes.

## Non-Goals
- Perfect performance prediction for every model/hardware combination.
- Replacing live canary, shadow inference, or SLO gates.
- Adding spot bidding or budget guardrails, which are already roadmap topics.

## Dependencies / Risks
- Requires accurate model resource profiles; missing profiles should degrade gracefully.
- Simulated results can become stale quickly in busy clusters.
- Needs clear UX to avoid implying guarantees the simulator cannot provide.

## Open Questions
- Should simulations be persisted for auditability when used before regulated rollouts?
- How should the simulator account for bursty stream FPS and quality changes?