Configure inference backend via compose up

[doringeman]

Added 3 new options: context-size, runtime-flags, and backend (which defaults to llama.cpp).
Also bumped docker/model-runner to docker/model-runner@6cf3f98.

This PR is for docker/model-runner#76.
You can test this by running this PR ⬆️ in a terminal with:

$ MODEL_RUNNER_PORT=8080 make run

And in a second terminal, build and test this PR with the following compose file:

services:
  model1:
    provider:
      type: model
      options:
        model: ai/smollm2
        context-size: 8192
        runtime-flags: "--no-prefill-assistant"
  model2:
    provider:
      type: model
      options:
        model: ai/llama3.2
        context-size: 1024

$ make install
$ MODEL_RUNNER_HOST=http://localhost:8080 docker compose --progress plain up
model1  Creating
model2  Creating
model2  Initializing model runner...
model2  Setting context size to 1024
model1  Initializing model runner...
model1  Setting context size to 8192
model1  Setting raw runtime flags to --no-prefill-assistant
model2  Successfully configured backend for model ai/llama3.2
model2  Created
model1  Successfully configured backend for model ai/smollm2
model1  Created

[xenoscopic]

LGTM. Do you know what's driving the large vendor directory change? Is it avoidable?

[xenoscopic]

Looks fine, just change paths and status if we decide to apply comments in docker/model-runner#76.

[doringeman]

Do you know what's driving the large vendor directory change? Is it avoidable?

It's due to the model-runner bump which brought in containerd (v2).