feat: Add EcoTune-based inference tuning module

[ust-xu]

• Adds a lightweight EcoTune core for inference-time hyperparameter optimization
• Keeps the contribution minimal: no benchmarks, demos, datasets, or docs in this PR
• Designed to be reused by future inference-tuning entrypoints and config-driven workflows
• Supports bootstrap from default decoding parameters

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull request overview

Introduces an EcoTune-based inference-time tuning core under flagscale.inference.tuning, implementing a small Bayesian-optimization style loop with a multi-fidelity GP surrogate and token-cost-aware acquisition to propose decoding/config suggestions under a budget.

Changes:

• Added multi-fidelity GP surrogate model for score prediction across config + fidelity.
• Added token-aware Expected Improvement acquisition and an optimizer implementing ask/tell with promotion to max fidelity.
• Added search space utilities and public package exports for flagscale.inference.tuning.ecotune.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
flagscale/inference/tuning/ecotune/surrogate.py	Implements a multi-fidelity GP surrogate with fit/predict and incumbent querying.
flagscale/inference/tuning/ecotune/search_space.py	Defines parameter dimensions, sampling, and config↔vector transforms.
flagscale/inference/tuning/ecotune/optimizer.py	Implements EcoTune optimizer loop (budgeting, suggestion, promotion, history).
flagscale/inference/tuning/ecotune/acquisition.py	Adds token-cost-aware Expected Improvement acquisition.
flagscale/inference/tuning/ecotune/init.py	Exposes EcoTune public API symbols.
flagscale/inference/tuning/init.py	Exposes tuning API from flagscale.inference.tuning.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.