chore: use new model runner syntax

[lionello]

Cf. https://docs.docker.com/ai/compose/models-and-compose/

Samples Checklist

✅ All good!

[defangdevs]

Rebased this branch onto current main (3ea9bfc), but defangdevs has no write access to this repo, so the rebased branch is parked at defangdevs/samples:lio/new-model-runner (commit 75acd1e). To update this PR:

git fetch https://github.com/defangdevs/samples.git lio/new-model-runner
git push --force-with-lease origin FETCH_HEAD:refs/heads/lio/new-model-runner

Conflict resolution: both mastra-extended hunks were dropped as superseded — main already converted that sample to the top-level models: syntax (cc5225a) and then switched it to portable aliases without x-defang-llm (3cd3c98), and the dev model on main is now gemma3 rather than the qwen2.5 this branch carried. The rebased commit touches only the two crewai files.

Pre-existing issues in the remaining crewai diff (in the PR as originally authored — I didn't change content during the rebase):

1. worker.depends_on still lists chat and embedding, which this PR removes as services — invalid compose.
2. Neither app nor worker is attached to the new top-level models (no per-service models: block), so CHAT_URL / CHAT_MODEL / EMBEDDING_URL / EMBEDDING_MODEL — which app/core/ai_clients.py and custom_llm.py read — are never injected. Compare mastra-extended on main, which uses endpoint_var / model_var attachments per service.
3. The chat model is renamed chat → llm in compose.yaml (but stays chat in compose.local.yaml), so even default injection would yield LLM_URL, not the CHAT_URL the app expects.
4. x-defang-llm: true is kept under models: here, while the samples already converted on main (mastra-extended, self-improving-mastra) omit it — worth deciding which is canonical.

[defangdevs]

Fixed the issues above in a follow-up commit on the parked branch (a0afb77c): the worker's stale depends_on entries are gone, both models are attached to the worker via endpoint_var/model_var (so CHAT_URL/CHAT_MODEL/EMBEDDING_URL/EMBEDDING_MODEL are injected again; compose.local.yaml inherits this through extends), the model is named chat in both files, and x-defang-llm is dropped to match the other converted samples.

Verified with defang compose config (built from defang main): both compose.yaml and compose.local.yaml render with the expected env vars on the worker.

Same command as before to update this PR (now brings both commits):

git fetch https://github.com/defangdevs/samples.git lio/new-model-runner
git push --force-with-lease origin FETCH_HEAD:refs/heads/lio/new-model-runner