v5: Minimal lazy-loading architecture for instant startup

Update Dockerfile to use app_v4 with background model loading

Add v4 with background model loading - prevents timeout by loading model after server starts

Add complete local Ollama setup with OpenELM - includes setup script, API server, test scripts, and documentation

Fix tokenizer issues: add sentencepiece dependency and robust loading strategy

Implement lazy loading - model loads on first request to avoid startup timeouts

Fix OpenELM tokenizer loading - use LlamaTokenizer as fallback

Add OpenAI API compatible endpoints for OpenELM models

Add Anthropic API compatible wrapper for OpenELM models

Add application files

initial commit