Add Qwen3-VL-Embedding-2B image input support

[cornmander]

Summary

• add a new Qwen3VLEmbedding API for Qwen3-VL multimodal embedding models
• support image inputs via embed_images(...) and embed_image_bytes(...)
• keep text support for Qwen3-VL via embed_texts(...) and existing Qwen3TextEmbedding
• add an internal Qwen3-VL vision module for image token embedding/injection
• update exports, feature wiring, README examples, and qwen3 tests

Implementation notes

• uses the qwen3 feature (candle backend) and enables dep:image for this feature
• image preprocessing follows Qwen3-VL patch/grid behavior (patch_size, merge_size, temporal duplication)
• prompt expansion replaces a single <|image_pad|> placeholder with the exact number of image patch tokens
• image embeddings are injected into the text stream before final hidden-state pooling

Validation

• cargo fmt --all
• cargo check
• cargo test --features qwen3 models::qwen3::tests -- --nocapture
• cargo test --features qwen3 --test qwen3 qwen3_vl_2b_text_embed -- --nocapture
• cargo test --features qwen3 --test qwen3 qwen3_vl_2b_image_embed -- --nocapture

New/updated interfaces

• Qwen3VLEmbedding::from_hf(...)
• Qwen3VLEmbedding::embed_texts(...)
• Qwen3VLEmbedding::embed_images(...)
• Qwen3VLEmbedding::embed_image_bytes(...)

[cornmander]

Pushed a follow-up formatting fix in 7894acd for the cargo fmt --all -- --check failure (import wrapping in src/models/qwen3.rs).

Current run status is action_required with no jobs started yet, so this PR likely needs a maintainer to approve workflow execution for the forked branch before CI can run again.

[Anush008]

Thanks for taking the time to contribute @cornmander

[Anush008]

Please add assertions equating embedding values from the Python counterpart code.

Like we do at https://github.com/Anush008/fastembed-rs/blob/main/tests/text-embeddings.rs

We have ensure Python and Rust produce the same vectors.

[cornmander]

Addressed the review feedback in a5c7c2f.

Changes made:

• Added Python-reference assertions in tests/qwen3.rs for qwen3_vl_2b_image_embed (embedding sums, first dims, and cosine).
• Aligned Rust VL image path with the official Python behavior:
  • Python-compatible ties-to-even image resize rounding.
  • MRoPE position-id construction for image tokens.
  • Interleaved MRoPE rotary application from mrope_section.
  • DeepStack visual feature injection into early decoder layers.

Validation run:

• RUN_QWEN3_VL_2B_IMAGE=1 cargo test --features qwen3 --test qwen3 qwen3_vl_2b_image_embed -- --nocapture
• RUN_QWEN3_VL_2B=1 cargo test --features qwen3 --test qwen3 qwen3_vl_2b_text_embed -- --nocapture

[Anush008]

Thanks @cornmander

[github-actions]

🎉 This PR is included in version 5.12.0 🎉

The release is available on:

• GitHub release
• v5.12.0

Your semantic-release bot 📦🚀