Spike at new CLI and SDK implementation

[gerred]

This reduces the code of the repeater into:

from typing import Any, Generator

from leapfrogai.llm import LLM, GenerationConfig


@LLM
class Repeater:
    def generate(
        self, prompt: str, config: GenerationConfig
    ) -> Generator[str, Any, Any]:
        yield prompt

That can be then run with:

leapfrogai main:Repeater

I might go down more of a FastAPI road where instead it's like:

from leapfrogai.backends import LLM, GenerationConfig

repeater = LLM()

@repeater.generate()
def generate(prompt: str, config: GenerationConfig) -> Generator[str, Any, Any]):
    yield prompt

That way it looks more like:

leapfrogai main:repeater

Would love thoughts on class vs. function style. This PR also completely separates the logic of serving the gRPC backend from the logic of generation.