feat(api): add non-generative task RPCs

[connorcarpenter15]

Summary

• add typed Embed, Classify, and Score RPCs to the existing OpenEngine service
• add common task inputs, dense and sparse embedding tensors, batch correlation, task usage, and grouped query/candidate scoring
• add per-model, per-task capability discovery for inputs, granularities, value semantics, normalization, limits, modalities, priority, and LoRA
• document batch-atomic lifecycle, validation, score semantics, and deterministic gateway-side reranking

Why

Generate cannot represent embeddings, classifier outputs, reward values, or model-native pair scores. Distributed frameworks therefore need engine-specific side channels for common non-generative inference.

This change establishes a portable typed surface across TensorRT-LLM, SGLang, and vLLM. It deliberately keeps reranking out of the engine protocol: gateways derive it from scalar Score results by stable-sorting, limiting, and reattaching gateway-owned documents.

Closes #10.

Validation

• buf build
• buf lint
• buf format --diff --exit-code
• buf breaking --against '.git#ref=origin/main'
• Protobuf descriptor compilation with protoc
• Markdown lint
• Documentation link check
• Commits include a DCO sign-off

Protocol checklist

• The canonical schema and API documentation are synchronized
• Presence, defaults, and invalid states are defined
• Acceptance, unary terminal, cancellation, and error behavior are defined
• Discovery and capability impacts are documented
• The change is additive against origin/main

Additional context

• Issue [Protocol]: Add portable non-generative inference task RPCs #10 was revised before implementation to retain Score and define Rerank as derived gateway/client behavior.
• Successful task batches are atomic and preserve request order while also returning presence-aware original indexes.
• Scalar, vector, and token-level scores cover native relevance/similarity, classifier/reward outputs, and causal label-token scoring.