Add lfm2 integration test

[TensorCruncher]

Description

Adds integration test for the recently added Lfm2 adapter. Modelled on the integration test file for NemotronH.

Let me know if I'm missing any tests / categories.

Do we need to add "LiquidAI/LFM2.5-230M" to the list of cached models for CI? Or is it okay to let it download each time since it's a relatively small model?

Related to #1510

Type of change

• New feature (non-breaking change which adds functionality)

[jlarson4]

Great tests @TensorCruncher! Thanks for putting this together.

The real 230M checkpoint is fine to keep, but we should mark it @pytest.mark.slow. This will allow it to run locally when I test PRs and when contributors run the tests before the submit, without bogging down CI with a minute+ download of a new model.

To cover this in CI without the big download, we'd typically use a tiny from_config test. You'll find a good example of that in test_nemotron_h_tiny.py. It builds a small model with random weights via AutoModelForCausalLM.from_config, so there's no Hub download and it runs in CI in a second or two. Could you add a test_lfm2_tiny.py alongside this one? Keep it small (a couple of layers so it covers both a conv and an attention layer) and it can carry the wiring, hook-coverage, and bridge-vs-HF forward checks.

Let me know if you have any questions!

[jlarson4]

The docs are pretty vague on this specific point, I am going to update them to clarify this pattern for future adapter creators. Thanks again for putting these together!