Add scheduler subsystem for gpu-remoting global orchestration

[Reggie1314129]

Summary

This PR adds a scheduler subsystem for gpu-remoting to support higher-level resource orchestration beyond the existing proxy/dispatcher workflow.

The new scheduler components introduce:

• resource models for jobs, GPUs, nodes, and clusters
• Redis-based message passing between scheduler and clients
• global scheduling across domains
• job-level resource allocation and reallocation logic
• experimental scheduling policies and preemption optimization
• workload simulation and profiling inputs for evaluation

What Changed

• add core scheduler data models:

  • gpu_info.py
  • job.py
  • node.py
  • cluster.py

• add scheduler runtime and policy modules:

  • resource_scheduler.py
  • resource_scheduler_dummy.py
  • sota_scheduler.py
  • util.py

• add control-plane and communication components:

  • msg_queue.py
  • global_scheduler.py

• add request handling and serving-side integration:

  • requese_handler.py

• add workload simulation and experiment assets:

  • client_simulator.py
  • cluster_client_sender.py
  • MOILP_latency_simu.py
  • job_info.csv

Motivation

The existing gpu-remoting flow mainly covers local GPU remoting and allocation. This PR extends the project with a scheduler layer that can:

• maintain structured cluster/job/GPU state
• coordinate allocation through Redis-backed messaging
• support multi-domain orchestration
• evaluate different scheduling and preemption strategies
• provide replay/simulation tools for scheduler experiments

Notes

• this PR includes both runtime scheduler components and experimental evaluation scripts in a single change set
• scheduler logic currently depends on the existing gpu-remoting configuration and Redis setup
• profiling / workload metadata is included to support scheduling decisions and experiments