GGUF export & llama.cpp

The GGUF button on the output block (or a run row in History) merges the model and writes a .gguf next to the run — ready for llama.cpp / Ollama / LM Studio:

• quant: f16 (exact, large), q8_0 (low error, lean — recommended) and K-quants like q4_k_m (needs llama-quantize on PATH or via TT_LLAMA_CPP).
• export progress is shown in the modal; on completion you get the file path and a download.
• metadata stays intact: tokenizer, architecture and attention biases are preserved — exported replies match the source model.