feat: add Gemma 3 4B ZeroGPU tab + /generate endpoint alongside molecular SLM (v2.0.0)

Fix Unicode subscript pIC50 regex: normalize ₅₀ → 50 before matching

fix: lazy model loading inside @spaces.GPU context for ZeroGPU compatibility

feat: add ZeroGPU support with @spaces.GPU decorator and spaces>=0.30.0

Fix 4 bugs: generic prompt, regex, float32, stub status

fix: load Qwen2.5-1.5B-Instruct base + LoRA adapter via PEFT

fix: add sentencepiece and tiktoken for Qwen2.5 tokenizer support

fix: replace ctransformers with transformers for Qwen2.5 safetensors loading

fix: replace llama-cpp-python with ctransformers for zero-compilation GGUF inference

Fix pIC50 output parsing from llama.cpp response

Add requirements.txt

Deploy Evolva inference API with graceful model-pending handling

initial commit