Add vector deduplication, indent context and litellm replacement

[ccmdi]

Summary

Vector-based deduplication has been added to prevent repeated flashcards by checking similarity against embeddings during generation. litellm was replaced with a lightweight custom client, and the Obsidian API integration was refactored to use JSONLogic instead of Dataview.

Changes

• Implemented a vector storage engine using JSON and API-based embeddings (Gemini/OpenAI) to index flashcard text.
• Added a multi-turn feedback loop where the AI proposes cards, receives similarity warnings based on the vector index, and revises content before submission.
• Replaced the litellm library with a custom call.py module that handles API requests for OpenAI, Anthropic, Google, and DeepSeek directly using httpx.
• Introduced vector CLI commands to index existing decks, check status, clear data, and test similarity matching manually.
• Refactored Obsidian note searching to use structured JsonLogic filters instead of concatenating DQL strings.
• Created an IndentedConsole class to manage output indentation scopes and added status spinners for long-running AI tasks.
• Updated the configuration to support vector deduplication settings and newer AI models like Claude Opus 4.5 and Gemini 3.
• Added conversation logging to save AI interaction history to the local configuration directory for debugging.
• Modified flashcard processing to automatically index new cards into the vector store upon creation.
• Updated unit tests to cover the new vector store, console wrapper, and search logic.