Add native reranking API for cross-encoder models

[leehack]

Motivation

llama.cpp has mature reranking support, including rank pooling, classifier metadata, named rerank chat templates, and server /rerank endpoints. The pinned b10333 native bindings already expose the required llama.cpp functions, but llamadart currently only exposes embedding-based similarity ranking.

Upstream: ggml-org/llama.cpp#9510

Proposed scope

• Add an explicit backend reranking capability.
• Add a public LlamaEngine.rerank(query, documents) API with typed ranked results.
• Use the model's named rerank template when available, with a compatible fallback prompt construction.
• Support batched document scoring with distinct sequence IDs.
• Handle both single-score and multi-class classifier output.
• Fail with a typed, actionable error for incompatible models/backends.
• Document supported platforms, expected model types, and scoring semantics.

This should initially be implemented in llamadart; b10333 appears to expose the required native symbols without a new llamadart-native release.

Validation

• Compare ordering and scores with the same model and inputs through upstream llama-server.
• Cover single-document and batched reranking.
• Cover multi-class classifier metadata where available.
• Cover long input/context-limit failure.
• Cover non-reranker and unsupported-backend rejection.
• Run a real BGE reranker smoke on representative native platforms.