[QDP Runtime] Add v1 State-Partitioned Distributed Runtime Draft

[400Ping]

Summary

This PR introduces a v1 draft of qdp-runtime, a state-partitioned distributed runtime skeleton for Mahout QDP.

The goal of this PR is to establish the core abstractions and control-plane scaffolding needed for multi-GPU and future multi-node execution. This is not intended to be a production-ready distributed runtime yet. Instead, it defines the v1 execution model, task lifecycle, placement policies, gather/reduce semantics, and runtime object tracking needed to iterate safely.

What This PR Adds

New crate

• qdp-runtime

Core runtime model

• state-partitioned distributed state metadata
• local/global qubit layout via PartitionLayout
• DistributedStateHandle and StatePartitionRef

Placement and topology

• RoundRobin, Weighted, and TopologyAware placement policies
• heterogeneous GPU-aware device capability model
• NVLink-aware topology metadata
• ClusterInventory and DeviceTopology

Coordinator / worker scaffolding

• worker registration
• in-process worker model
• coordinator job planning
• partition task generation
• partition-to-worker mapping

Task lifecycle

• Pending / Assigned / Running / Completed / Failed
• retry policy
• lease timeout skeleton
• task result reporting

Output handling

• runtime object/output registry
• gather planning via GatherPlan
• metric reduction planning via ReducePlan
• host-side metric aggregation for Sum / Mean / Min / Max / Concat

Examples and docs

• qdp-runtime/examples/local_runtime_smoke.rs
• qdp-runtime/examples/local_runtime_benchmark.rs
• qdp/docs/runtime/RUNTIME_V1.md

Scope of This PR

This PR focuses on the v1 runtime draft and control-plane design.

It intentionally does not attempt to provide:

• full multi-node transport
• persistent GPU object store
• partition migration
• dynamic repartitioning
• production-ready fault tolerance
• complete benchmark suite integration

Design Notes

The current v1 draft is centered around a state-partitioned execution model.

That means:

• a logical state may be partitioned across devices
• partition layout is currently contiguous amplitude blocks
• placement can be weighted and topology-aware
• gather and metric reduction are explicit runtime operations
• runtime outputs are tracked via a coordinator-side object registry

This PR also keeps NVTX instrumentation hooks in place so that future runtime bottlenecks can be profiled more easily.

Current Limitations

• transport is still in-process / metadata-oriented
• local executor support is minimal
• examples are smoke/benchmark scaffolding, not full workflow integration
• topology support is advisory metadata in v1
• object tracking exists, but a persistent GPU-resident object store is future work

Follow-Up Work

• wire the runtime into a real single-node local execution path end-to-end
• add persistent runtime object storage semantics
• improve retry/reassignment behavior
• extend topology-aware gather/reduce logic
• add benchmark integration with existing QDP benchmark workflows
• add multi-node transport and worker communication

Related Issues

Related to #1210

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[400Ping]

cc @viiccwen