Improve robustness of get_task HF pipeline invocations

[vblagoje]

Is your feature request related to a problem? Please describe.
As a Haystack user, I noticed that when HF infra is temp down I can't use a newly added Cohere model. The issue has been traced to the PromptModelInvocationLayer.supports invocation order. Because HFLocalInvocationLayer is listed in haystack/nodes/prompt/invocation_layer/__init__.py before CohereInvocationLayer, HFLocalInvocationLayer gets registered via __init_subclass__ mechanism before CohereInvocationLayer. Then when we invoke HFLocalInvocationLayer.supports, the get_task call blocks indefinitely, thus preventing the use of the Cohere-based PromptNode.

Describe the solution you'd like
I would like Haystack PromptNode initialization to be as robust as possible and not dependent on HF infra outage, especially when I am not using one of the HF invocation layers for PromptNode.

Describe alternatives you've considered
No resolution is possible atm aside from waiting for the outage to stop

Additional context
Add any other context or screenshots about the feature request here.