Rethink testing strategy for HF-based components

[anakin87]

In Haystack 2.x, most components based on Hugging Face libraries don't have integration tests.

Why?
The integration tests were avoided mainly to keep the CI fast:

• some of these tests would require downloading models and performing slow inference
• others (for TGI and TEI) would require running docker services

The problem
We want the CI to be fast but integration tests are useful for detecting changes and bugs.
(We run some e2e tests that also cover some of these components, but I think they serve a different purpose and are not sufficient)

The plan
So I would propose that we rethink our strategy for testing these components.
I don't have strong ideas about that. Perhaps nightly integration tests?