ci(mesh): add Atomesh accuracy and benchmark workflows

[wanzhenchn]

Motivation

This PR adds dedicated Atomesh CI coverage for mesh correctness, mocker-based PD routing performance, and dashboard visibility.

Atomesh has different validation needs from the full ATOM model CI. The goal is to keep mesh CI focused, lightweight, and actionable:

• Validate Atomesh standalone entrypoint integration through a representative accuracy subset.
• Continuously benchmark PD separation routing behavior with mocker-based workloads.
• Publish benchmark history to a dedicated dashboard for performance tracking.
• Avoid running unrelated ATOM, vLLM, and SGLang CI when a PR only changes mesh-specific code.

Summary

Atomesh accuracy validation

• Adds Atomesh standalone accuracy validation using USE_ATOMESH_ENTRYPOINTS=1.
• Keeps a small representative model subset with local levels to the Atomesh workflow, so shared models_accuracy.json ownership remains unchanged.
  • pr: Meta-Llama-3-8B-Instruct
  • main: DeepSeek-R1-0528
  • nightly: DeepSeek-V4-Pro MTP, gpt-oss-120b
• The workflow expands coverage by trigger type:
  • Pull request runs only pr models for fast entrypoint validation.
  • Push to main runs pr + main models.
  • Scheduled and manual runs execute pr + main + nightly models.
• Publishes Atomesh accuracy results into the existing benchmark dashboard data path for dashboard reuse.

This keeps Atomesh standalone validation focused on USE_ATOMESH_ENTRYPOINTS=1 integration coverage, while the full ATOM accuracy workflows continue to own the complete model matrix.

Atomesh mocker benchmark

• Adds a mocker benchmark workflow focused on PD separation routing scenarios.
• Runs representative topology coverage:
  • 1P1D
  • 2P1D
  • 3P1D
• Runs each topology across consumer concurrency levels:
  • 1, 2, 4, 8, 16
• Records request throughput, latency metrics, failed requests, request count, duration, and CI run metadata.
• Aggregates per-cell results into a markdown summary and benchmark-action compatible JSON.

Mocker benchmark dashboard

Live: https://rocm.github.io/ATOM/atomesh-mocker-dashboard/

• Adds a custom Atomesh mocker benchmark dashboard.
• Provides Performance, Accuracy, and Trends views.
• Performance view includes:
  • Request throughput chart
  • Combined latency chart for avg / p99 / p999 latency
  • Detailed performance table with configuration, concurrency, request count, commit, and CI run link
  • Download JSON support
• Accuracy view reuses Atomesh accuracy dashboard data and displays GSM8K validation results.
• Trends view shows historical benchmark changes by configuration and concurrency.
• Publishes the custom dashboard HTML together with benchmark data to gh-pages.

CI routing optimization

• Adds mesh-only PR skip rules for unrelated CI workflows.
• Mesh-specific changes no longer trigger:
  • ATOM Test
  • ATOM vLLM Test
  • ATOM SGLang Test
• Atomesh-specific workflows still run for mesh source, Atomesh workflow, Atomesh script, and Atomesh dashboard changes.
• Mixed PRs that touch mesh plus ATOM/vLLM/SGLang paths still preserve full relevant CI coverage.