Split CI benchmarks into PR, scheduled, and manual tiers

[parthdongre]

Summary

Restructure performance CI so ordinary development gets fast regression feedback while very large physical stress tests run on an intentional cadence.

Why

FrameVitals has strong benchmark coverage, including billion-cell-scale workloads, but the largest workflows are too expensive to treat like normal per-push CI. Separating tiers will reduce noise and cost without weakening performance validation.

Proposed tiers

1. PR guardrail: small deterministic performance regressions with short runtime.
2. Scheduled benchmark: broader representative workloads on a daily/weekly cadence.
3. Manual stress: extreme physical-scale runs through workflow_dispatch before major releases or architecture changes.

Acceptance criteria

• PR benchmark jobs have a documented runtime budget.
• Extreme multi-billion-cell workloads do not run on every normal development push.
• Scheduled benchmarks retain representative native/fallback validation.
• Manual stress workflows remain easy to run before releases.
• Benchmark artifacts/results remain reproducible and attributable to a commit.
• Documentation explains which benchmark tier should be used for each kind of change.

Priority: P0 — CI efficiency / performance
Suggested labels: performance, ci, infrastructure