refactor(vlm): add owned host prefill preprocessing (#859)

[inureyes]

Summary

• Add a backend-neutral owned prepared-prefill DTO with validated embeddings, positions, additive attention bias, logical token history, and modality metadata.
• Add a LLaVA host preprocessor and canonical filtered loader that retain only the vision tower, projector, and text embedding weights while bypassing transformations not applied by IREE.
• Share exact image-placeholder cardinality and overflow-safe expansion semantics with the existing MLX path, and export evaluated contiguous tensors through the fallible FFI boundary.
• Provide a deterministic checkpoint-free fake plus focused parity, dtype, shape, capacity, and loader-filter tests.

Validation

• cargo fmt --all -- --check
• cargo check -p mlxcel-core --lib --tests
• cargo check -p mlxcel --lib --tests
• cargo test -q -p mlxcel-core multimodal::tests
• cargo test -q -p mlxcel multimodal::host_preprocessor::tests --lib
• cargo test -q -p mlxcel multimodal::vlm_prompt::tests --lib
• cargo test -q -p mlxcel loading::vlm::llava::tests --lib
• cargo test -q -p mlxcel vision::merge::tests --lib
• cargo clippy -p mlxcel-core --lib --tests -- -D warnings
• cargo clippy -p mlxcel --lib --tests -- -D warnings

Closes #859