Add support for Qwen3Guard-Stream model with OVModelForGuard

[p-wysocki]

What does this PR do?

• Qwen3Guard-Stream is an Qwen3 decoder with LM head replaced with four token classification heads
• Existing feature-extraction doesn't work, because OVModelForFeatureExtraction.forward() only reads last hidden state
• Introduced OVModelForGuard as an elegant solution - the alternative would be to significantly complicate OVModelForFeatureExtraction
• The branch runs the actual model correctly for both stateless and stateful pipelines

Note

• This is my first contribution to optimum-intel, so I'm open to any discussion and feedback

Ticket

• CVS-191209

[p-wysocki]

@rkazants @popovaan could you please review?