[Refactor & fix] Request: refactor and fix swe-agent example

Refactor:

The current swe-agent example (examples/experimental/swe-agent) is based on nemo-gym and nv-mini-swe-agent. This was a temporary solution with internal considerations. However, this solution is not clean when provided as an example for miles. We may not need to use nemo-gym for server setup, but directly use mini-swe-agent as a lightweight external module for simplicity, which is more like other examples.

Fix:

(If you choose to refactor, you may not need to follow the current design and fix it in this way. But just to show the problem for reference)

Also, the current nv-mini-swe-agent does not apply chat templates to sglang engine. It (sglang_model.py) currently uses /generate, but should be changed to OpenAI APIs instead. Also, the processing in generate_with_swe_agent.py needs to be modified along with this change.

Request:

Overall, I suggest doing a full refactor of this example. If anyone has time, feel free to look at it. Also, I'm happy to chat at anytime to discuss or help.