Llama-cpp-cffi

Fixed:
- Stop condition in `Model.completions`

Fixed:
- Stop condition in `Model.completions`

Fixed:
- In `server.py`, lock is now used correctly

Added:
- `CompletionsOptions.stop` is optional strgin which triggers stop of generation

Changed:
- In `llama_cpp.py`, use `threading.Lock` instead of `DummyLock`
- In `server.py`, use `asyncio.Lock` instead of `threading.Lock`

Changed:
- Updated all requirements

Changed:
- Reimplemented `is_cuda_available` and `is_vulkan_available`