Mlipy

Fixed:
- Docker environment without GPUs.

Changed:
- Do not import `server.py` from package by default.

Added:
- Detect `AMD` and `NVIDIA` GPUs because `llama.cpp` needs to know valid arguments.

Changed:
- Default port is `4000` instead of `5000`.

Added:
- `client` special `model_id='echo/echo'` which returns same text back without going to server.
- examples/defs.py has new grammar templates.

Changed:
- `server` uses `web.RouteTableDef` for defining routes.

Added:
- `amazon/MistralLite` 7B model demos.
- `google/gemma` 7B and 2B models demos.

Deprecated:
- `candle` models in `README.md`.
- `candle` instructions in `README.md`.

Fixed:
- `llama.cpp`/`candle` engine does not handle wrong/missing model.
- Early check if engine one of `llama.cpp` or `candle`.