compile llama-cpp-python with -j1 for low-mem build

add requests (gradio cli imports it)

install llama-cpp-python 0.3.19 wheel directly (no compile)

switch to Docker SDK + abetlen prebuilt llama-cpp-python wheel

switch to transformers + merged fp16 (avoid llama-cpp build OOM)

fix: pin python 3.11 (llama-cpp-python wheels available, avoids py3.13 build)

fix: use prebuilt CPU wheels for llama-cpp-python (avoid OOM on builder)

fix: bump llama-cpp-python 0.3.2 -> 0.3.20 for Qwen3.5 GGUF support

fix: bump gradio 4.44 -> 5.9 + audioop-lts for py3.13

initial demo: Qwen3.5-0.8B LoRA q4_k_m GGUF on CPU

initial commit