[Feature]: Add Qwen3-Embedding model support (0.6B/4B/8B) for LLM API encode_only mode

🚀 The feature, motivation and pitch

Native support for the Qwen3-Embedding model family (0.6B, 4B, 8B) for use with the LLM API's encode_only mode.

Problem

The Qwen3 embedding models don't work correctly out of the box with encode_only. TensorRT-LLM returns logits (post lm_head), whereas the correct output should be the last prompt token's hidden state before the lm_head (model card's transformers usage)

Proposed support

Add a new architecture/model class, e.g. Qwen3ForEmbedding, for use with LLM(..., encode_only=True).

The model class would:

• Reuse the existing Qwen3 transformer backbone without applying lm_head.
• Return last prompt token's hidden states from LLM.encode(...) instead of logits.
• Leave normalization and MRL/custom-dimension slicing to the caller or serving layer.

Alternatives

No response

Additional context

Models: Qwen3-Embedding-0.6B, Qwen3-Embedding-4B, Qwen3-Embedding-8B

I'm happy to contribute a PR for this feature

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.