feat(router): direct native-gRPC dispatch beside the request plane

[ishandhanani]

Phase 2 of the runtime-optional frontend roadmap (contract 3, dispatch). Stacked on #18.

Summary

A second RoutingHost transport: admitted KV-routed requests go straight to a worker's native engine RPC when the worker advertises one, so the sidecar leaves the request path while discovery still resolves membership and endpoints. Workers that do not advertise keep using the request plane, so a fleet can migrate one worker at a time.

• lib/llm/src/kv_router/direct_dispatch.rs: DirectEngine (an AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<LLMEngineOutput>>>, the exact shape the request-plane client returns), DirectEngineFactory, DirectDispatchRegistry (worker id -> engine) with a feeder on the runtime-config watch: connect on runtime_data["native_grpc_endpoint"], drop on departure or endpoint change, retry failed connects on the next membership change. install_direct_engine_factory is the process-wide seam so lib/llm never depends on an engine protocol crate (the sidecar crates depend on lib/llm via backend-common, so the dependency must point this way).
• RoutingHost::with_direct_dispatch; dispatch_selection (routing_host/kv.rs) chooses the direct engine when the selected worker has one. Everything around it is untouched: guard.abort() still releases the booking before the error propagates (contract 1), the DynamoError chain migration::is_migratable inspects is preserved, and cancellation reaches the engine through the request context. Documented gap: PushRouter fault detection does not see direct-dispatch failures; liveness comes from discovery.
• backend-common: direct_engine(Arc<dyn LLMEngine>, mode) exposes the worker-side EngineAdapter, so a frontend gets identical streaming, deferred decode abort (decode must reach the engine so NIXL can release transferred KV), and error semantics as the request-plane path. Re-exports the direct-dispatch types and ModelRuntimeConfig.
• vLLM sidecar: --advertise-grpc-endpoint / DYN_ADVERTISE_GRPC_ENDPOINT publishes native_grpc_endpoint + native_grpc_mode in the worker's runtime data. VllmSidecarEngine::connect_direct (async discovery + start) and VllmDirectEngineFactory implement the factory over the existing gRPC client, request conversion (convert.rs), and status_to_dynamo error mapping (connect/timeout/cancelled/unavailable -> the same BackendErrors the migration taxonomy already classifies).
• Assembly: DYN_ROUTER_DIRECT_DISPATCH=vllm in the Python frontend installs the vLLM factory; entrypoint/input/common.rs attaches a registry to every KV routing host when a factory is installed. Bindings gain the dynamo-vllm-sidecar dependency.
• Docs: sidecar-backends.md (Direct Dispatch section), frontend-configuration.mdx, vLLM sidecar README.

Validation

• cargo test -p dynamo-llm --lib -- kv_router: 334 passed. New direct_dispatch_serves_registered_workers_and_releases_on_failure: a registered fake engine serves the request and the booking is released on completion; a failing fake releases the booking and surfaces ErrorType::CannotConnect through the error chain; an unregistered worker never touches the fake.
• cargo test -p dynamo-vllm-mocker -p dynamo-vllm-sidecar: 4 + 23 + 13 passed. New direct_engine_factory_streams_from_the_mock_server: the factory connects to the mock vLLM gRPC server, streams 3 tokens with a Length finish, leaves no active requests, and an unreachable endpoint fails to connect.
• cargo clippy --all-targets on dynamo-llm, dynamo-backend-common, dynamo-vllm-sidecar, dynamo-vllm-mocker: clean. cargo check of the Python bindings: clean. cargo fmt and docs_lint.py clean.
• Not yet run: an end-to-end sidecar-backed deployment with DYN_ROUTER_DIRECT_DISPATCH=vllm against a real vLLM gRPC server and the worker-shutdown migration tests (needs a GPU worker); these are the Phase 2 exit gates.

[ishandhanani]

Closed while Stack 1 (#45 to #53) and Stack 2 (#40) go upstream; the branch stays for later stacks (direct dispatch, catalog leases, runtime features). See #33.