feat: add 'none' LLM provider for chunk-only storage mode

[nicoloboschi]

Summary

• Adds a none LLM provider option for running Hindsight as a chunk store with semantic search but without any LLM dependency
• Replaces the workaround of setting HINDSIGHT_API_LLM_PROVIDER=mock for chunks-only mode
• When provider is none: retain auto-uses chunks mode, recall works normally, reflect returns HTTP 400, consolidation/observations are disabled

Changes

• providers/none_llm.py (new): NoneLLM provider class + LLMNotAvailableError exception
• config.py: Forces retain_extraction_mode=chunks and enable_observations=false when provider is none
• llm_wrapper.py: Registers none as a valid provider (no API key required)
• memory_engine.py: Guards on reflect, consolidation, mental model refresh; forces chunks mode in per-bank config
• http.py: Catches LLMNotAvailableError → HTTP 400 on reflect and mental model refresh
• configuration.md: Documents the none provider with usage example
• test_none_llm_provider.py (new): 12 tests covering unit, config, integration, and HTTP API behavior

Test plan

• NoneLLM.call() and call_with_tools() raise LLMNotAvailableError
• NoneLLM.verify_connection() succeeds (no-op)
• Config validation forces chunks mode and disables observations
• Retain works (stores chunks without LLM calls)
• Recall works (semantic search via embeddings)
• Reflect raises LLMNotAvailableError at engine level
• Consolidation is skipped with {skipped: true}
• Mental model refresh raises LLMNotAvailableError
• HTTP reflect endpoint returns 400
• HTTP retain/recall endpoints work normally