Collect plugin process metrics on demand

[popduke]

Plugin instance CPU and resident memory are missing from the Runner's metrics endpoint, so users need separate instrumentation and collectors to observe plugins. This change includes those observations in GET /metrics through request-driven Runner → Pipeline → Plugin SDK collection.

• Add a private metrics stream alongside Plugin Control. Java and Rust SDKs handle sampling, OpenTelemetry encoding, filtering and reconnection without an author-facing metrics API or periodic sampling.
• Define shared CPU/RSS semantics and launch identity. The first CPU sample establishes a baseline; failed, unavailable and timed-out observations remain missing rather than becoming zero. Collection stays independent of plugin lifecycle and data delivery.
• Keep queue measurements owned by the Pipeline and attribute submission/completion queues to the Source instance and egress queues to the actual Sink instance.
• Organize reusable IPC and process-metrics packages under sdk/<language>, with shared IPC, process-metrics and Plugin SDK contracts. Update build paths, scaffolds, tests and documentation, including the Event-Driven Stateful Transducer description of Tenon Lua.

Runner, Pipeline and plugins built with the updated SDKs must be upgraded together. This change does not add compatibility negotiation for older SDKs.

Validation:

• Complete macOS arm64 Rust/Java verification, including generated plugins, cross-language process contracts, shutdown and standalone package consumption.
• Native Linux arm64 sampler, metrics-stream and Java/Rust process checks. Two Java lifecycle tests initially hit startup timeouts; isolated reruns and the subsequent full suite passed without timeout or production-code changes. The initial cause remains undetermined.
• Current branch deployed on macOS arm64 and Linux amd64 with rebuilt MQTT/Kafka plugins. Verified CPU/RSS and queue metrics through the API and CPU/RSS in Console, sustained a local MQTT → Kafka → MQTT round trip at 10 messages/s, and received all 10 remote MQTT → Kafka canary messages.
• Public documentation, license-header checks and git diff --check passed during implementation; the complete PR diff also passes git diff --check.