feat: support estimated recommendations via Config Explorer integration

[anfredette]

Summary

Integrate Config Explorer's capabilities into the recommendation pipeline to provide estimated recommendations when benchmark data is unavailable for a model/GPU combination.

Approach

• Use capacity_planner.py to determine which GPU configurations can physically fit a model (memory feasibility filtering)
• Use gpu_recommender (BentoML roofline model) to generate synthetic performance estimates for configurations lacking benchmark data
• Feed these estimates into the existing scoring/ranking pipeline as a fallback

User flow

1. User identifies target GPUs and models
2. System checks all models and GPUs against GPU and Model catalogs
3. For combinations without benchmark data, use capacity planning + roofline estimation
4. Present estimated recommendations alongside benchmark-backed ones

Reference

See migration plan PR #129 Follow-on Step 3 and meeting notes (Mar 30, 2026).