Add parametrized steerable model tests (Transformers backend)

[kudos07]

Summary

This draft PR introduces a parametrized test suite for steerable models
covering multiple Hugging Face models under the Transformers backend.

It focuses on one aspect of issue #1717:

• Expands initialization and inference coverage to multiple models.
• Runs each inference several times to surface flaky behavior.
• Adds a shared Hugging Face cache fixture to prevent rate-limit errors.
• Includes a backend parameter to allow future vLLM/llama_cpp coverage.

Notes

• vLLM tests are skipped on Windows; CI can re-enable them.
• This is a draft PR as discussed with @RobinPicard to align on test organization before expanding further.