feat(studio): add preference-driven Agent quality management

[evanlowe]

Adds a Quality management entry in Agent details, starting from user-defined evaluation preferences. Users can review and adjust model-generated dataset and evaluator preferences before generating the two components together. Existing Evaluations and feedback remain unchanged.

• Add independent, selectable suggestions to all ten preference text areas, with shared Agent context preparation, caching, cancellation and field-level retry
• Generate 50–300 evaluation cases in bounded parallel batches, validate counts and duplicate inputs, and display the cases in a drawer
• Generate overall, tool capability and Skill evaluators concurrently, with vertically stacked cards, loading placeholders, dimension accordions and Agent-specific scoring rationales and prompts
• Ground generation in available topology, root and child instructions, tools, Skills, knowledge bases, memory and environment metadata, preserving ownership and information gaps
• Share the Agent form generation model for Volcengine and BytePlus, use structured output with thinking disabled, and return complete captured provider diagnostics with credentials redacted
• Add preference-linked component filtering, outdated indicators, partial-success recovery, input validation, bilingual copy and matching WebUI assets

Preferences, datasets and evaluator definitions currently live in page memory. Persistence, evaluation execution and result analysis are outside this change; the results tab shows an unavailable state.

Validation:

• CI passed: Python 3.10 and 3.12 unit suites, frontend/backend gates, release bundle verification, license checks, pre-commit and secret scanning
• Synced with current origin/main before submission
• Frontend: all 1,235 tests passed; production build, packaged-asset verification and locale consistency checks passed
• Backend: 171 targeted tests passed, 3 skipped, including both providers, authorization, schemas, parallel generation, cancellation and redacted diagnostics
• Pre-commit across all files passed, including Ruff and secret scanning
• Pyright passed for the quality backend, new tests and shared planner model changes; the existing CLI file retains the same 33 diagnostics reproduced on main, with no new diagnostics
• Browser checks covered suggestions, manual edits, partial results, retry and narrow layouts; live generation checks covered recommendations and a 50-case dataset

Additional startup checks passed 32 tests, with test_complete_studio_app_defers_model_catalog_network_runtime failing locally. The same failure was reproduced on an unchanged checkout of main.