feat(iterative-research): pre-flight validations and runtime model failure handling

[skorphil]

Summary of Changes

This pull request introduces critical, user-facing inline validation checks and error rendering to prevent silent or confusing failures inside the iterative-research plugin.

1. Pre-flight Validation & Halting:

   • Validates that at least one search engine (SEARXNG_URL or TAVILY_API_KEY) is configured in the valves.
   • Validates that a backend MODEL is specified in the valves.
   • If validations fail, streams a highly visible markdown warning banner at the top of the response, logs the failure, and halts execution before any loops or LLM operations take place.

2. Runtime LLM Failure Handling:

   • Catch empty/failed responses from background LLM planning and synthesis tasks.
   • Rather than silently assuming "All Gaps Resolved" on empty/failed completions, halts processing immediately, closes any open <thinking> tags cleanly, and renders a clear runtime failure banner.

3. Removed Auto-detection:

   • Removed model auto-detection block (Models.get_all_models()) entirely to strictly enforce explicit configuration and prevent unexpected lookups or model resolutions.

4. Testing & QA:

   • Cleaned up obsolete tests and added robust test coverage for validation rules: test_pipe_validation_no_search_engine, test_pipe_validation_no_model, test_pipe_validation_only_tavily_configured, test_pipe_validation_only_searxng_configured, test_pipe_validation_valves_none_and_empty_values.
   • Added planning and synthesis failure checks: test_pipe_planning_llm_failure, test_pipe_synthesis_llm_failure.