refactor(api): remove Drain RPC, consolidate proto files, trim EngineError

[connorcarpenter15]

Summary

Follow-up to #27. Two related contract cleanups in one PR:

1. Remove the Control.Drain RPC (+ DrainRequest/DrainResponse/DrainState).
2. Consolidate the proto file set 10 → 8 and trim dead message/field surface.

Package name, message names, and field numbers are unchanged, so the consolidation is wire-compatible and transparent to whole-package codegen — only which file a message lives in changes.

1. Drain removal

Graceful shutdown doesn't need an engine-facing drain RPC. The Dynamo sidecar (the single OpenEngine client) can drain without it: stop admitting (discovery unregister), let in-flight Generate streams complete, poll GetLoad for quiescence, and cancel stragglers via gRPC stream cancellation into the engine's existing Abort path.

Dropped lifecycle.proto's now-unused error.proto import. Kept the drain-adjacent state signals that stand on their own — HEALTH_STATE_DRAINING, KvConnectorInfo.supports_drain (validated by the sidecar for disagg connectors), and ERROR_CODE_DRAINING — since an engine still drains on its own SIGTERM.

2. Consolidation

• observability.proto → server.proto. After the runtime-events removal it held only load reporting (GetLoadRequest/LoadInfo/RankLoadInfo) — "observability" had become a misnomer. Load is the dynamic half of what server.proto already reports (static identity/capacity/roles).
• input.proto → generation.proto, its sole remaining consumer. TokenIds/Modality/MediaItem move as-is.

3. Dead-code / field trim

• Task-input types (TaskInput, MultimodalTaskInput, TaskInputType) — orphaned when Embed/Classify/Score were removed. Deleted.
• EngineError fields — dropped details (never set or read) and retry_after_ms (only ever absent), plus the now-unused struct import. Kept retryable: all three servers populate it (SGLang computes it from a 503) even though the sidecar doesn't yet read it — removing it would discard a real signal, not dead code.

Result

Control surface: GetServerInfo · GetModelInfo · GetLoad · Health · Abort · LoadLora · UnloadLora · ListLoras · GetKvEventSources · SubscribeKvEvents (+ Inference.Generate).

File set 10 → 8: openengine · generation (+input) · kv · model · server (+observability/load) · lifecycle · lora · error.

Validation

• protoc -I proto --experimental_allow_proto3_optional builds the full descriptor set clean.
• markdownlint-cli2 clean on all Markdown.

Consumer follow-up (separate repos, not this PR)

• Drain: the sidecar's drain() override and the three servers' Drain handlers are removed; LLMEngine::drain reverts to the default no-op and shutdown relies on the is_quiescent/GetLoad + grace-period + stream-cancellation path already in Worker.
• EngineError: servers stop setting details / retry_after_ms (no reader exists, so nothing breaks); retryable handling is unchanged.

Compatibility

Wire-breaking for the removed RPC and the two EngineError fields (which no consumer reads). The file consolidation itself changes no wire format.