Performance & Bug Fix: ModelPool Optimization and VLM Detection Improvements

Changes

• Performance: Unified VLM detection logic to eliminate redundant getContainer() calls
• Bug Fix: Added Gemma DWQ special handling - quantized models lose vision capabilities and should be treated as LLM
• Feature: Added extraEOSTokens support for QAT tokenization bugs (e.g., <end_of_turn> in Gemma models)
• Feature: Add Gemma 3n - Text Only (LM) support
• Optimization: Cached model type detection to avoid repeated registry lookups

Testing

• ✅ All 44 unit tests pass
• ✅ Release build completes successfully (107.26s)
• ✅ CI validation completed

Impact

• Faster model loading through reduced redundant operations
• Correct model type detection for quantized variants
• Better tokenization handling for models with EOS token bugs